ANNUAL REPORT 2020
CONTENTS
6 Executive Summary
8 Organization
10 ZIB Structure
12 Research to Fight COVID-19
14 New ZIB law passed by the parliament of Berlin
16 Economic Situation in 2020
22 Understanding and Modeling Complex Biological Systems
36 Wind under Your Wings
48 NHR@ZIB: Taking the Next Step in HPC
54 Simple Explanations, Sparsity, and Conditional Gradients
64 Research Campus MODAL: Success Stories
78 References 3 81 Publications
90 Imprint
ZIB.DE
EXECUTIVE SUMMARY
The corona crisis is massively changing the lives of individuals and their coexistence. This is true for everyday life as well as for the life at a research institution like ZIB. Many things have changed since the beginning of 2020. We learned how to work and do research together via virtual platforms and had to change the way we interact, coordinate, and more generally em- brace digitalization. How fundamental and sustainable are these developments? Will research and scientific exchange be different after the corona crisis? At ZIB, cohesion was great and we believe that we kept the special team spirit of the institute, across all the different groups and people. However, at the same time everybody is eagerly waiting for the pandemic situ- Executive Summary ation to improve such that we can meet each other in person again.
Research related to the corona crisis. Soon after the National alliance for High-Performance Computing. In outbreak of the pandemic, several research projects re- 2020, the German federal and state governments decided lated to corona started at ZIB, with model-based as well to jointly initiate and found a National High-Performance as data-based approaches. Our main project, entitled Computing Alliance (NHR). The NHR alliance will create a “Model-based investigation of school closures and other nationally coordinated network of high-performance com- mitigation measures for COVID-19” brought together sever- puting centers. Eight centers, one of which is ZIB, were al research groups from TU Berlin, the Robert Koch Institute, selected through a national competition, evaluating both and ZIB. The overarching goal of this joint project – fund- their scientific as well as HPC competency. The evaluation ed by the German Ministry of Research and Education process involved both a review by the German Research (BMBF) from 2020 until 2023 – is to investigate the effects of Foundation (DFG) and an independent strategy committee non-pharmaceutical interventions on the infection dynamics appointed by the Germany’s Joint Science Conference. The of COVID-19, ranging from school closures, over econom- NHR center at ZIB will receive more than EUR 70 million of ic and societal lockdown to vaccination. The project also funding for the next ten years; see also the feature article aims to improve our understanding of infection chains and on the NHR center. the dynamics of spread within urban areas such as, e.g., Berlin, as well as in regional and national contexts. The re- Amendment to the ZIB Act. On the basis of an intensive sults attracted attention of the federal government, among debate with relevant parties lasting more than a year, an others, and were widely discussed and commented on in amendment to the ZIB Act was formulated, discussed in the media. the responsible bodies and passed by the Parliament of the State of Berlin on December 2, 2020. The new law together with institutions like ZIB. In “NHR@ZIB: Taking the defines ZIB as a research institute for scientific computing, Next Step in HPC,” further insight is provided about the application-oriented mathematics, and high-performance creation of the NHR center at ZIB and what it means for the computing with an interface to artificial intelligence. It also future of high-performance computing at ZIB. The feature entangles ZIB more closely with the partners of the Berlin articles “Understanding and Modeling Complex Biological University Alliance. More details can be found in the article Systems” reports on ZIB’s recent work on seamlessly inte- “New ZIB law passed by the parliament of Berlin.” grating modeling approaches and machine learning for enabling new ways for extracting “dynamical laws” from Research Campus MODAL enters its second funding the observation of complex systems on a macroscopic or phase. After successful completion of its first funding phase microscopic level. This overview is complemented by the from 2014 to 2020, the Research Campus MODAL was eval- feature article “Simple Explanations, Sparsity, and Condi- uated by an expert review panel and the grand jury of tional Gradients” which explains why it is often important 7 the German Ministry of Research and Education (BMBF). to find sparse solutions to optimization problems that ap- MODAL’s proposal for a second phase was granted for pear at the core of machine learning methods and how the funding period 2020-2025. The grant includes funding this sparsity requirement can be realized algorithmically. Fi- of EUR 10 million by the BMBF matched by a considerably nally, “Wind under your Wings” reports on recent progress larger amount of the participating private companies. regarding discrete-continuous free flight planning related MODAL hosts more than 60 researchers from academia to the development of hybrid approaches combining dis- and industry under a single umbrella. More details, espe- crete optimization algorithms with optimal control methods cially some success stories of public-private partnerships in order to compute globally optimal free flight trajectories at ZIB, can be found in the feature article on the Research with high accuracy in an efficient way. Campus MODAL. Despite all the problems resulting from the COVID pan- More insights. In addition to the topics already men- demic, 2020 was a very successful year for ZIB with key tioned, this annual report provides insights into a variety of successes that will positively impact on ZIB’s development other success stories and gives a general overview of ZIB’s for a long time. Many other very positive developments organization and key factors for its successful develop- make us confident that our institute has a bright future. ment. For example, the feature article “Research Campus Put differently, ZIB continues to be a place for excellent MODAL: Success Stories” outlines what can be achieved research and first-rate scientific services and infrastructure. regarding research transfer if private partners closely work ORGANIZATION
The Statutes
The Statutes, adopted by the Board of Directors at its meeting on June 30, 2005, define the functions and procedures of ZIB’s bodies, determine ZIB’s research and development mission and its service tasks, and frame upon the composition of the Scientific Advisory Board and its role.
Organization Administrative Bodies
The bodies of ZIB are the President and the Board of Thorsten Steinmann Directors (Verwaltungsrat). Der Regierende Bürgermeister von Berlin, Senatskanzlei – Wissenschaft und Forschung President of ZIB Prof. Dr. Christof Schütte Dr. Jürgen Varnhorn Senatsverwaltung für Wirtschaft, Energie und Betriebe Vice President Prof. Dr. Sebastian Pokutta Prof. Dr. Manfred Hennecke Bundesanstalt für Materialforschung und -prüfung (BAM) The Board of Directors was composed in 2020 as follows: Thomas Frederking Prof. Dr. Peter Frensch Helmholtz-Zentrum Berlin für Materialien und Energie Vice President, Humboldt-Universität zu Berlin (Chairman) (HZB)
Prof. Dr. Christian Thomsen Prof. Dr. Heike Graßmann President, Technische Universität Berlin (Vice Chairman) Max-Delbrück-Centrum für Molekulare Medizin (MDC)
Prof. Dr. Günter Ziegler The Board of Directors met on June 3, President, Freie Universität Berlin 2020, and November 16, 2020. Scientific Advisory Board
The Scientific Advisory Board advises ZIB on scientific and technical issues, supports ZIB’s work, and facilitates ZIB’s cooperation and partnership with universities, research institutions, and industry.
The Board of Directors appointed the following Prof. Dr. Rolf Krause members to the Scientific Advisory Board: Université della svizzera italiana, Lugano, Switzerland
Prof. Dr. Jörg-Rüdiger Sack (Chairman) Ludger D. Sax Carleton University, Ottawa, Canada Grid Optimization Europe GmbH
Prof. Dr. Frauke Liers Prof. Dr. Reinhard Schneider Friedrich-Alexander-Universität Erlangen-Nürnberg, Université du Luxembourg, Luxembourg Germany Prof. Dr. Dorothea Wagner Prof. Dr. Michael Dellnitz Karlsruher Institut für Technologie (KIT), Karlsruhe, Germany Universität Paderborn, Germany (Membership shall be held in abeyance during the continu-
ing service as Chair of the German Science Council) 9
The Scientific Advisory Board met on July 2, 2020, online.
BOARD OF DIRECTORS
SCIENTIFIC ADVISORY BOARD Chairman: Prof. Dr. Peter Frensch Humboldt-Universität zu Berlin (HUB)
PRESIDENT Prof. Dr. Christof Schütte
VICE PRESIDENT Prof. Dr. Sebastian Pokutta
MATHEMATICS FOR MATHEMATICAL PARALLEL AND ADMINISTRATION LIFE AND MATERI- OPTIMIZATION AND DISTRIBUTED COMPUTING ALS SCIENCES SCIENTIFIC INFORMATION
Prof. Dr. Christof Schütte Prof. Dr. Sebastian Pokutta Prof. Dr. Alexander Reinefeld Dr. Kathrin Rost-Drese (acting) (until August 1, 2020) ZIB STRUCTURE
MATHEMATICS FOR LIFE MATHEMATICAL AND MATERIALS SCIENCES ALGORITHMIC INTELLIGENCE
C. Schütte S. Pokutta
MODELING AND VISUAL AND AI IN SOCIETY, SCIENCE, APPLIED ALGORITHMIC NETWORK SIMULATION OF DATA-CENTRIC AND TECHNOLOGY (AIS²T) INTELLIGENCE OPTIMIZATION COMPLEX PROCESSES COMPUTING METHODS (A²IM)
M. Weiser T. Conrad S. Pokutta T. Koch R. Borndörfer
COMPUTATIONAL VISUAL DATA MATHEMATICAL PRESCRIPTIVE HEALTH ANATOMY AND ANALYSIS OPTIMIZATION SYSTEMS LOGISTICS PHYSIOLOGY METHODS
M. Weiser, S. Zachow D. Daum, H. Hege A. Gleixner J. Zittel G. Sagnol
COMPUTATIONAL GEOMETRIC DATA INTERACTIVE PREDICTIVE INDIVIDUAL MOLECULAR DESIGN ANALYSIS AND OPTIMIZATION AND METHODS TRAFFIC PROCESSING LEARNING (IOL) (BAG)
ZIB Structure M. Weber C. von Tycowicz S. Pokutta M. Petkovic E. Swarat
COMPUTATIONAL COMPUTATIONAL FAIRNESS, ACCOUNTABILITY, MULTI MOBILITYLAB NANO OPTICS DIAGNOSIS AND TRANSPARENCY, AND ENERGY THERAPY PLANNING EXPLAINABILITY (FATE) SYSTEMS
S. Burger S. Zachow S. Pokutta J. Zittel N. Lindner
COMPUTATIONAL BIOINFORMATICS SYSTEMS BIOLOGY IN MEDICINE SUSTAINABLE ENERGY PLANNING S. Winkelmann T. Conrad C. Tawfik
UNCERTAINTY EXPLAINABLE AI QUANTIFICATION FOR BIOLOGY
M. Weiser, T. Sullivan V. Sunkara
COMPUTATIONAL HUMANITIES
N. Conrad
RESEARCH SERVICE UNIT IT RESEARCH SERVICE UNIT DIGITAL DATA AND INFORMATION AND DATA SERVICES FOR SOCIETY, SCIENCE, AND CULTURE (D²IS²C)
C. Schäuble B. Rusch, T. Koch
BRAIN BERLIN RESEARCH PROCESS MODELING AND DATA MANAGEMENT DIGIS: RESEARCH AND KOBV LIBRARY AREA INFORMATION MANAGEMENT SERVICES COMPETENCE CENTER NETWORK NETWORK DIGITALIZATION BERLIN BERLIN-BRANDENBURG
C. Schäuble C. Schäuble, K. Rost-Drese S. Prohaska A. Müller B. Rusch, S. Lohrum PARALLEL AND ADMINISTRATION DISTRIBUTED COMPUTING
A. Reinefeld K. Rost-Drese (acting)
DISTRIBUTED SUPERCOMPUTING ALGORITHMS
F. Schintke T. Steinke
DISTRIBUTED HPC DATA CONSULTING ZIB is structured into four divi- MANAGEMENT sions: three scientific divisions F. Schintke T. Steinke and ZIB’s administration.
SCALABLE HPC ALGORITHMS SYSTEMS Each of the scientific divisions is composed of two departments F. Schintke C. Schimmel that are further subdivided into re- search groups (darker bluish col-
MASSIVELY ALGORITHMS FOR 11 PARALLEL INNOVATIVE or) and research service groups DATA ANALYSIS ARCHITECTURES (lighter bluish color). F. Schintke T. Steinke
DYNAMICS OF COMPLEX MATERIALS
F. Höfling, T. Kramer
RESEARCH SERVICE UNIT DIGITAL DATA AND INFORMATION LEGEND FOR SOCIETY, SCIENCE, AND CULTURE (D²IS²C) Scientific divisions B. Rusch, T. Koch and departments
Research groups FRIEDRICH-ALTHOFF- DIGITAL OPEN SCIENCE AND ZIB LIBRARY KONSORTIUM PRESERVATION RESEARCH DATA Research service groups (OSARD) Research service unit U. Kaminsky W. Peters-Kottig W. Dalitz K. Lachmann RESEARCH TO FIGHT COVID-19
Research to Fight The COVID-19 virus has caused a worldwide pandemic with more than 120 million positive cases and more than 2.5 million deaths by March 2021. As long as effective medical treatment and vaccination are not available everywhere, non-pharmaceuti- cal interventions such as social distancing, self-isolation, and quarantine as well as far-reaching shutdowns of economic activity and public life are the only available strate-
COVID gies to prevent the virus from spreading. Researchers at ZIB contributed to mathematical models and simulation tools that are capable of predicting the spread of the infection
-19 for different combinations of non-pharmaceutical interventions. Reports based on these tools supported governmental policy decisions on regional and federal level.
In 2020, in a worldwide effort, researchers started examined and model-based policy recommendations to construct a multitude of mathematical models for are formulated. predicting the spread of the infection. At ZIB, we con- sidered model-based and data-based approaches. The main objective of MODUS-COVID is to build a mi- Our main project entitled “Model-Based Investigation cro-model in the form of an agent-based model (ABM), of School Closures and Other Mitigation Measures for together with a pipeline that allows the simulation of COVID-19” (MODUS-COVID) brought together several a synthetic population with realistic movement pat- research groups from the TU Berlin, Robert Koch In- terns, subjection of the synthetic individuals to infection stitute, and ZIB. The overall goal of this joint project dynamics, and then testing the response of infection (funded by the German Ministry of Research and Ed- dynamics to different interventions to subsequently ucation (BMBF) from 2020 until 2023) is to investigate evaluate the effectiveness of these interventions. The the effects of non-pharmaceutical interventions on the main ABM allows the population of Berlin (approxi- infection dynamics of SARS-CoV-2, initially with a spe- mately 5 million people, 3.6 of which are living inside cial focus on schools closing. The project also aims to the city limits) to be described moving about during the improve the understanding of infection chains and the day, meeting other agents at work, in school, on public dynamics of spread within urban areas as well as in a transport, at leisure activities, or when shopping. The regional and nationwide context. In addition to examin- mobility side of the model had been worked on for ing schools closing, other measures and combinations many years before the outbreak of SARS-CoV-2 and is of measures to curb the spread of SARS-CoV-2 are based on cell-phone data [1]. The ABM while others, such as social and economic costs, assumes that call for weaker countermeasures. Therefore, finding an infection may the optimal compromise of countermeasures re- occur when an infectious quires the solution of a multi-objective optimization and a susceptible agent are in the same building problem that is based on accurate prediction of or vehicle simultaneously. The transport model future spread of infection for all combinations of used is activity-based, meaning that it contains countermeasures under consideration. A strategy for complete daily activity chains for every agent. construction and solution of such a multi-objective
These activities have types (for example home, optimization problem with real-world applicability 13 work, leisure, shopping), durations, and locations. was presented in [2]. The strategy is based on com- The activity chains are available for a typical week- plementing the ABM by a surrogate macro-model day, a typical Saturday, and a typical Sunday. This that is much less computationally expensive and means that for every simulated week, the weekday can therefore be used in the core of a numerical model is run five times and then the Saturday and solver for the optimization problem. The resulting Sunday models. In these simulations, every time two set of optimal compromises between countermea- agents meet and one of them is contagious and the sures (Pareto front) can be computed, and then be other susceptible, the probability of an infection is discussed as a background for policy decisions. computed. This results in a large-scale spatiotempo- ral model, simulations of which require utilization of The ABM developed in MODUS-COVID can also be a high-performance computing infrastructure. All ABM used to model the spread of the B.1.1.7 mutation simulations in the context of MODUS-COVID were and other variants, for investigating the widespread performed on the supercomputer operated by ZIB. use of rapid tests, and the effects of vaccination strategies. It allows quantitative predictions for ABM models provide microscale simulations of the differentiated strategy building blocks. The ABM- spread of infection for different intervention strate- based tool for simulation-based investigation of in- gies. However, their very high computational effort fection dynamics developed as part of the project prevents the use of ABMs as the prediction core of a is available under an open-source license and will multi-objective optimization scheme that is required be available free of charge following the project. for identifying the best intervention strategies: in- The biweekly MODUS-COVID reports that were ad- terventions have to meet conflicting requirements dressed to the German Ministry of Research and where some objectives, such as the minimization Education (and have also been part of the report- of disease-related deaths or the impact on health ing to the federal government) can be found on systems, demand for stronger countermeasures, www.zib.de. NEW ZIB LAW PASSED BY THE PARLIAMENT OF BERLIN New ZIB Law passed by the Parliament of Berlin
On the basis of an intensive debate with relevant parties lasting more than a year, an amendment to the ZIB Act was formulated, discussed in the relevant bodies and passed by the Parliament of the State of Berlin on December 2, 2020. The new law links ZIB more closely with the partners of the Berlin University Alliance.
ZIB now is defined as a research institute of the State In order to create the right basis for this mission, the of Berlin with the purpose to promote science and new law reconstitutes ZIB’s Board of Directors with a research in the field of scientific computing, appli- different structure. From 2021 on, the Board of Direc- cation-oriented mathematics and high-performance tors is composed of the responsible member of the computing, including related development and re- government of the State of Berlin, the heads of the search services. partner institutions of the Berlin University Alliance (BUA) – the presidents of Berlin’s three largest uni- The tasks of ZIB encompass the development of versities, FU, HU, and TU Berlin, the Chief Executive models and algorithms, in order to use computer Officer of Charité – as well as one of the directors simulations and optimization methods as well as da- of the Helmholtz Zentrum Berlin. This composition re- ta-driven procedures to solve problems from natural flects the importance of the link between ZIB and the sciences, engineering, life sciences and medicine as BUA partners, and, by this, the importance of ZIB for well as the social sciences and humanities. Research the future of Berlin’s research environment. is complemented by related development and re- search services, in particular regarding the opera- In addition, the management of ZIB will be expand- tion of high performance research infrastructures. ZIB ed and strengthened. The new law creates the posi- works in close cooperation with the universities of the tion of an additional vice-president. This position will state of Berlin and the Charité – Universitätsmedizin be filled through a joint professorship with one of Berlin, Berlin’s university hospital, in order to meet the BUA’s partner institutions, thus further strengthening research service requirements of the parties involved. the connection to BUA. Moreover, the head of ZIB’s NEW ZIB LAW PASSED BY THE PARLIAMENT OF BERLIN 15
administration will take over the role of the director of finance, so that the scientific management can focus more on strategic and scien- tific matters as well as on ZIB’s cooperation with the international scientific community.
Furthermore, ZIB is given the right to independently enter contracts with scientific and technical personnel, which means a significant simplifica- tion of administrative procedures and a strengthening of strategic options regarding recruitment of personnel and fostering junior researchers. Last but not least, ZIB obtains more freedom regarding its economic situation. This leeway is necessary in order to be able to act in a more professional and targeted manner in the area of research-related services in the future. However, it also means that ZIB has to reform parts of its activities, e.g., regarding the expansion and digitization of its financial accounting.
In conclusion, the new law brings quite some changes for ZIB, its members and its environment. It will need some time to implement all these changes appropriately but the result will be an even stronger and more agile institute. ECONOMIC SITUATION IN 2020
In 2020, the total income of ZIB comprised 30.9 million euros. The main part of this was made available by the Federal State of Berlin as the basic financial stock of ZIB (19.3 million euros) including investments and Berlin’s part of the budget of HLRN at ZIB. The second largest part of the budget resulted from third-party funds (7.0 million euros) acquired by ZIB from public funding agencies (mainly DFG and BMBF) and via
Economic Situation industrial research projects. This was complemented by a variety of further grants, such as the HLRN budget made available by other German states or the research service budget of KOBV, summing up to almost 4.6 million euros in total.
ZIB INCOME
€19,281,330 €7,059,810 62% 23% Core Budget by Third-Party State of Berlin Funds
€4,560,630
15% Further Grants The Zuse Institute Berlin (ZIB) finances its scientific to 5.219 million euros in 2020, industrial third-party work via three main sources: the basic financial stock projects also declined from 2.450 million euros to of the Federal State of Berlin and third-party funds 1.840 million euros. In total, 7.060 million euros in from public sponsors and those of industrial cooper- third-party funding marked a significant decrease, ation contracts. primarily due to the transition from the first to the second funding period of the Research Campus In 2020, ZIB raised third-party funding through a MODAL and the resulting funding gap, as well as large number of projects. Project-related public third- reduced expenditures in the area of travel, guests, party funds declined from 6.136 million euros in 2019 and conferences. 17
ZIB THIRD-PARTY FUNDS BY SOURCE
€1,446,110
€2.030.200
€1,840,330
€1,743,170
20% BMBF 26% 29% 23% incl. FC Modal Industry DFG Other Public Funds ZIB THIRD-PARTY FUNDS IN EUROS
€10,000,000
€8,000,000
€6,000,000 Economic Situation
€4,000,000
€2,000,000
2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
Industry Public Funds €10,000,000
€8,000,000
€6,000,000 19
€4,000,000
€2,000,000
2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
Industry Public Funds Spin-Offs
Computing in Technology GmbH (CIT) Lenné 3D GmbH 1992 | www.cit-wulkow.de 2005 | www.lenne3d.com Mathematical modeling and development of 3-D landscape visualization, software develop- numerical software for technical chemistry ment, and services
RISK-CONSULTING JCMwave GmbH Prof. Dr. Weyer GmbH 2005 | www.jcmwave.com 1994 | www.risk-consulting.de Simulation software for optical components Database marketing for insurance companies onScale solutions GmbH Intranetz GmbH 2006 | www.onscale.de 1996 | www.intranetz.de Software development, consulting, and services Software development for logistics, database for parallel and distributed storage and com- publishing, and e-government puting systems
Visage Imaging GmbH Laubwerk GmbH (Originating from the ZIB spin-off Visual 2009 | www.laubwerk.com Concepts GmbH) Construction of digital plant models 1999 | www.visageimaging.com Advanced visualization solutions for diagnostic 1000shapes GmbH imaging 2010 | www.1000shapes.com
Spin-Offs Statistical shape analysis atesio GmbH 2000 | www.atesio.de Quobyte Inc. Development of software and consulting for 2013 I www.quobyte.com planning, configuration, and optimization of Quobyte develops carrier-grade storage soft- telecommunication networks ware that runs on off-the-shelf hardware
bit-side GmbH Keylight GmbH 2000 2015 I www.keylight.de Telecommunication applications and visualiza- Keylight develops scalable real-time Web ser- tion vices and intuitive apps. The focus is on prox- imity marketing, iBeacon, and Eddystone for Dres. Löbel, Borndörfer & Weider GbR / interactive business models LBW Optimization GmbH 2000 | www.lbw-optimization.com Optimization and consulting in public transport LBW Optimization GmbH was founded in 2017 and is a spin-off of LBW GbR Number of Employees
In 2020, 219 people were employed at ZIB; of these, 134 positions were financed by third-party funds. The number of employees grew in comparison to 2020, mainly because of increase in funding by the State of Berlin and as a consequence of the start of the second funding period of the Research Campus MODAL.
1/1/2020 1/1/2021
MANAGEMENT 3 0 3* 2 0 2* *without temporary management 21 21 78 99 28 86 114 SCIENTISTS
41 0 41 41 3 44 SERVICE PERSONNEL
15 1 16 14 1 15 KOBV HEADQUARTERS
0 40 40 0 44 44 STUDENTS
80 119 199 85 134 219 Total Total Total Temporary Temporary Permanent Permanent ANTHROPO EV OL UNDERSTANDING UTIONARCENTURY
EC LOGY VE ONOMIC Y OUPLINGS UISITIONS OL CH SIMPLE ACQ APPR AND MODELING C Y EV HUMAN MA DIFFERENTIAL AO ION THEMOA CHEMISTRTAT LINE AT OGY S ME IC CH BIOL OMPUCHANISTI PROC AL C RELA C SIMULA AR COMPLEX CHINE AUTOMATA MA LIBRE UNPR TIONSH TION ES LEARNING SOCIOLOGY AL EDICTABLY IP S Y HISTORIC S Marcus Weber | HOLOG A APPRO BIOLOGICAL PSYC FINITION XIMATED C SCIENCE SILIENCY DE FIELD PHENOMENRE JOURNALS ONFER [email protected] SYSTEMS
ENCE XS LEER TIONS MPEVE CENT CONESS BR RANDOM Y PRECISEL RTURBA S | +49-30-8 41 85-189 INT ABSCENCE PE BEHAVIOR ELLIGENCE CONCEPT C OMPLE EARLY S EVOLUTION SYSTEMACTIONPHYSICAL SICS G DE X THEORY PHY TE SCIENCES ROBOTICS EVOLVE OUPLIN RMINISTI INFORMATION EFINED C NARRATIVESNETWORKUNDPROBLE SEARCHE Complex systems are characterized Researchers at ZIB have CONSTR MOD RE CTIV by their multiple scales. To better contributed to many topics AININ EMERGENT TICS
C M OLLE understand and ultimately allow the in this area, such as: G GOVERMENTS COMPLE CCYBERNE control of such complex systems, COMPLEX ELLINCHALLENGEDSTRIVEN insights about their underlying laws · Data Analysis: “Understanding Bio- OMPUTIONALTYNEUROSCIENCELIFEFORMALIZED are needed that can then be used to logical Data: How Machine Learning C ATTRACTOR COMPLEXI SIS X simulate those systems for validation Can Help” Y G PREDICTION and hypothesis testing. Researchers · Simulation: “Speeding up Simula- Y ANALHIST MAIN VIOR at ZIB develop new mathematical tions Using Artificial Intelligence” FULL DE ARTIFICIAL G USE RA DYNAMICAOTIC BEHA methods needed for these tasks. In · Complexity Reduction: “Reducing CHALLENGEDTIONALIVELO OR CH M UILBRIU particular, new approaches based the Complexity: a Data-Driven Ap- FR Y EQ on time-series analysis and machine proach for Molecular Systems” TY DER learning have been researched that · Modeling and Optimization: “Learn- AC PED OR enable new ways for extracting “nat- ing Dynamical Laws from Noisy Data” TALS ural laws” from the observation of complex systems on a macroscopic REMARKIN or microscopic level. ANTHROPO EV OL UTIONARCENTURY
EC LOGY VE ONOMIC Y OUPLINGS UISITIONS OL CH SIMPLE ACQ APPR C Y EV HUMAN MA DIFFERENTIAL AO ION THEMOA CHEMISTRTAT LINE AT OGY S ME IC CH BIOL OMPUCHANISTI PROC AL C RELA C SIMULA AR CHINE AUTOMATA MA LIBRE UNPR TIONSH TION ES LEARNING SOCIOLOGY AL EDICTABLY IP S Y HISTORIC S HOLOG A APPRO PSYC FINITION XIMATED C SCIENCE SILIENCY DE FIELD PHENOMENRE JOURNALS ONFER
ENCE XS LEER TIONS MPEVE CENT NESS BR
CO 23 RANDOM Y PRECISEL RTURBA S INT ABSCENCE PE BEHAVIOR ELLIGENCE CONCEPT C OMPLE EARLY S EVOLUTION SYSTEMACTIONPHYSICAL SICS G DE X THEORY PHY TERMINISTISCIENCES ROBOTICS EVOLVE OUPLIN INFORMATION EFINED C NARRATIVESNETWORKUNDPROBLE SEARCHE CONSTR MOD RE CTIV AININ EMERGENT TICS
C M OLLE G GOVERMENTS COMPLE CCYBERNE COMPLEX ELLINCHALLENGEDSTRIVEN COMPUTIONALTYNEUROSCIENCELIFEFORMALIZED ATTR COMPLEXI X ACTOR YSIS G PREDICTION Y ANAL MAIN VIOR FULL DE HIST ARTIFICIAL USE RA TIC BEHA G CHALLENGEDTIONALIVELO ORDYCHNAMICAO M UILBRIU FR Y EQ AC TY PED ORDER TALS
REMARKIN Understanding and Modeling Complex Biological Systems
Understanding Biological Data: How Machine Learning Can Help Biomedical experiments often lead to large data collections. This could be images from high-resolution microscopy devices, genomic sequences from tumors, or a list of all protein concentrations that are contained in blood samples. Understanding the data and extracting useful information manually is not only tedious but even becomes impossible if the size of the data gets too large. Computer-based algorithms can help to automate and speed up this process. However, one of the main challenges is to create methods that recognize the useful information and ignore the rest. Machine learning-based approaches are a way to achieve this by training an algorithm until it delivers the desired output. Researches at ZIB have contributed to the design of such methods for several applications. 25 Within the area of machine learning, researchers are build- of machine learning algorithms. Most of the time, unfor- ing and applying algorithms that can improve with experi- tunately, both options are true, and, on top of that, the ence. While this is quite a normal part of our everyday lives available data contains errors and noise. In this situation, – given that we are humans – it is still surprising to see that solutions are needed that somehow allow the pretraining a machine can also do the same thing, at least in some of the machine learning algorithms on synthetically gen- specialized areas, such as object and pattern recognition erated data first, where the quality can be controlled. or prediction of future events. In some of these areas, many However, this seems to be a paradox: if the system that machines (or algorithms) already outperform humans not should be learned is known, one does not need to learn only in accuracy but also in terms of speed – owing to the it. But if the system is unknown, how can data about it be ever-increasing power of new processor technologies. generated? Researchers at ZIB contributed to overcoming this challenge for applications in biomedical data analy- The two main types of machine learning are supervised sis. The key insight is that it is beneficial to pretrain – or
Understanding and Modeling and unsupervised learning. Supervised learning is focused initialize – a machine learning algorithm with data that is Complex Biological Systems on figuring out the relationship between the given data similar – but not exactly the same – to the target data. This and their respective labels, for example, whether an ECG data can be synthetic, e.g., resulting from simulations using measurement represents a medical condition known as a well-established (but not perfect) model for the system bradycardia. In this case, the algorithm does not know at hand. This allows the algorithm to learn the general a priori how to recognize bradycardia – instead, it will structure and is then followed by a fine-tuning step using learn this from the data. Once this relationship has been the real data. This combination leads to better results com- learned, the algorithm can suggest (or predict) the label of pared to only training on the available real data and has an unknown instance and with this, might support a medi- been demonstrated to work very well for proteomics and
cal doctor recognizing heart diseases [Weimann2021]. The heart activity (ECG) data [Iravani2019, Weimann2021]. main promise of machine learning is: the more data has been used to train such an algorithm, the better the quality One of the main goals while analyzing complex systems in of the predicted labels. general and biological systems in particular, is to under- stand the underlying rules or laws of the dynamic behavior The other main type of machine learning are unsupervised over time. Opposed to the previously described situation, algorithms. Here, the data does not have labels. The goal in this case the question is not about predicting labels is to find patterns in the data that can then be used to of an unknown instance (supervised learning) or finding identify sub-groups (or clusters) in the data. In biomedical (sub-)groups in large datasets (unsupervised learning), applications, such cluster-identification algorithms help re- but the idea is to get insights into the mechanisms that searchers to better understand biological subtypes based can be used to derive a mathematical formulation about on their genomics profile, which could lead to a better a system. Once this can be formulated, simulations can understanding of the underlying mechanisms [Rams2020]. be performed and deeper analysis and hypothesis testing can be carried out. ZIB researchers have presented new Although more and more biological data is generated approaches that could successfully demonstrate the appli- these days, still, in most situations, either not enough data cability of these ideas to real-world systems from the life or not enough labels are available for a successful training sciences [Melnyk2020, Zhang2019]. Speeding up Simulations Using Artificial Intelligence
Simulations of biological systems play a crucial role in many biomedical applications, such as drug development. Here, for example, one is interested whether a drug is actually able to bind to a receptor long enough to really trigger the desired action, such as pain relief. Unfortunately, detailed and long enough simulations of these kinds of molecular processes are still infeasible these days. Researchers at ZIB have developed new approaches that can circumvent this problem by using methods from artificial intelligence.
Detailed simulations of a biological system – sition path theory. The committor function is such as the interaction between a drug and given by a receptor – offer a deep understanding of the processes involved, for example, the time needed for the drug to really trigger the de- L χ = 0, χ = 1 in the target state B, sired function or the on- and off-rates for the χ = 0 in the initial state A, drug to bind to the receptor. However, in most where L denotes the infinitesimal generator cases these simulations are computationally of the molecular dynamics (MD) of the re-
too costly and cannot be performed even on ceptor, drug, and their environment. In the 27 a supercomputer. To overcome this problem, simplest case where MD is modeled by an researchers at ZIB have developed alterna- overdamped Langevin equation, the infinites- tive approaches that require only a fraction imal generator L takes the form of a partial of the computation time and still deliver the differential operator, which turns the above required information. equation into a partial differential equation (PDE). For more general situations, see [Do- The approach taken is based on so-called nati2020]. Similar constructions can be used partial differential equations and committor to compute the off-rates, for example, the fre- functions. Their solution provides all the es- quency of the event that the drug leaves the sential information about the macroscopic binding pocket of the receptor, or ligand-pro- behavior of the studied stochastic dynamic tein dissociation rates [Ray2020], leading to microsystem. For each possible state of a mi- linear equations of the form crosystem, the key is that a committor function indicates how likely this state is to evolve to
a certain target state B, before it returns to 1 2 L χ = Ɛ χ − Ɛ (1 − χ), the initial state A. By this, for example, the the solution of which combines committor on-rate with which a drug (starting in the un- functions and mean holding times for appro- bound state A) binds to the binding pocket priate boundary conditions, where and
(state B) of the receptor can be computed denote small scalar numbers related1 to the2 Ɛ Ɛ from the committor function by means of tran- rates, see [Ernst2020]. Solving PDEs with Artificial Intelligence
The solution of these kinds of PDEs would allow the calculation of the desired macroscopic properties of the full microscopic system. Unfortunately, solving these equations can still be hard or even impossible due the extremely high dimension. Understanding and Modeling Complex Biological Systems
Figure 1: The flowchart of the algorithm ISOKANN, which combines adapted molecular simulations with deep learning. Figure 2: After just two iteration cycles of ISOKANN, the neural network has distinguished between two states of the microsystem and provides the relevant information about the macrosystem.
To tackle this, a new approach was found that turns the original problem of 29 solving the PDE into a so-called functional fixed-point problem. This can be solved by a fixed-point iteration that turns out to be computationally feasible. The developed method, named ISOKANN [Rabben2020], is based on an artificial neural network that is trained to represent the iterated function (see Figure 1). ISOKANN proved to be highly efficient in cases that could not be solved by alternative means. The key ingredient is that the solution of the PDE can be regarded as a continuous classification problem: whether a mi- crostate can be interpreted as “the drug is still at the receptor,” or not (see Figure 2). This classification is done very efficiently by the neural network and allows application of adaptivity during the learning phase, where the set of training points can be adjusted in order to put an emphasis onto the “interesting” part of the molecular process (i.e., onto the transition region of the molecular system). Reducing the Complexity: A Data-Driven Approach for Molecular Systems
Ab initio molecular dynamics allows simulations on spatial and temporal resolutions that are not accessible by traditional experimental methods. Often however, one is not primarily interested in the atomistic detail of a molecular system, but more in the chemically and biologically relevant processes that emerge on longer time and length scales. Gaining insight into these processes requires the careful construction of physically consistent reduced models from simulation data. Central here is the identification of good reaction coordinates (i.e., the “slow” degrees of freedom that form the parameter space of the reduced model). Understanding and Modeling Complex Biological Systems
A GEOMETRICAL VIEW OF REACTION COORDINATES
For given reaction coordinates (RCs), the cor- This viewpoint allows a nonparametric repre- responding reduced dynamical law can be sentation of the RCs to be learned by applying
constructed by truncation techniques such as unsupervised manifold learning methods, such the Mori-Zwanzig formalism. However, the ac- as the diffusion-maps algorithm, to an empirical tual consistency with the microscopic dynamics approximation of the system’s transition densities. will depend heavily on the initial choice of the This requires simulation data in the form of many RCs. ZIB researchers were involved in formulat- short, independent trajectories, whose gener- ing a new, general characterization of optimal ation is favored by modern massively parallel timescale-preserving RCs [Bittracher2020a]. It computation hardware. Furthermore, the manifold revolves around the property of the system’s tran- learning step can be enhanced by embedding sition probabilities that with increasing lag time, the data into a reproducing kernel Hilbert space their dependence on the quickly equilibrating, prior to this, which regularizes the problem and unimportant degrees of freedom vanish, so that minimizes the number of free parameters [Bittra- eventually they essentially depend only on the cher2020b]. The result is a robust and flexible relevant slow degrees of freedom. machine learning algorithm that has been used to identify RCs, transition pathways, and metasta- Geometrically, this means that the system’s tran- ble conformations of high-dimensional molecular sition densities, considered as points in a certain systems such as peptides and proteins (Figure 3). metric function space, evolve toward a low-di- mensional manifold. A parametrization of that manifold corresponds to the desired RCs. D B
0.05 B 0 ξ 3 A C − 0.05 A − 0.04 C 0 − 0.1 2 0.03 ξ 0 0.04 D − 0.03 ξ1 Figure 1: Three-dimensional reaction coordinates evaluated at sampled configurations of the Alanine dipeptide. The edges (arrows) in the low-dimensional manifold structure correspond to transition pathways between local minima of the potential energy surface.
MODEL REDUCTION OF 31 SPARSELY SAMPLED SYSTEMS
Another important aspect of reducible systems 40.81 is that, due to the existence of an underlying re- 40.81 duced model, the microscopic dynamics cannot become arbitrarily complex. This has the effect that the amount of dynamical data required to approximate the reduced dynamics up to a fixed level of accuracy does in fact not depend on the dimension of the full system, but only on the di- Latitude 40.7 −74.01 −73.94 mension of the reduced system. For the subclass
1 P of discrete Markov chains, ZIB researchers and Aggregate 1 P^ Aggregate 2 colleagues were able to formulate a comprehen-
Aggregate 3 k λ 0.5 sive theory behind this phenomenon, and explain Aggregate 4 40.7 why accurate reduced Markov jump models of −74.01 − 73.94 0 high-dimensional metastable systems can be Longitude 1234 k constructed from vastly sparse trajectory data [Bittracher2021]. This theory also has applica- tions beyond molecular dynamics; for example, Figure 4: Four-state aggregated Markov model of the taxi traf- fic on Manhattan Island (left). The aggregates were identified they demonstrated how an accurate large-scale based on traffic measurements at only a small subset of all model for urban taxi traffic can be constructed microscopic states (top right, red boxes). The full and reduced model show excellent agreement (in the form of leading from only a small number randomly-chosen mea- eigenvalues of the respective transition matrices, bottom right) surement points (Figure 4). Learning Dynamical Laws from Noisy Data
Besides classification and prediction, a recent trend in machine learning is to infer the underlying physical laws from data that generated them, which makes use of data not only for concrete application purposes but also for advancing the understanding of complex processes. Since physical laws can usually be formulated as simple algebraic expressions, one way is to look for “simple” dynamical laws predicting the observed data well. Understanding and Modeling Complex Biological Systems LEARNING DYNAMICAL LAWS
A tractable form of the symbolic regression ap- [Gelß2019]), combinations with the Mori-Zwan- proach is to build linear combinations of a few zig formalism, and Taken’s embedding theorem atoms selected from a large dictionary of po- leading to extensions of SINDy with memory tential contributions. This is particularly attrac- terms like in sparse identification of nonlinear tive for inferring dynamical laws (i.e., right-hand autoregressive models (SINAR) [Wulkow2020], sides in ordinary differential equation systems with essential contributions by ZIB researchers.
such as (bio)chemical reactions) due to their uniform structure. Algorithms such as orthogo- In another branch of the literature, especial- nal matching pursuit (selecting promising terms ly Koopman operator theory and extended sequentially) and formulations such as basis dynamics mode decomposition (EDMD), ap- pursuit (regression with a sparsity-enhancing proximations of dynamical systems by transfer L1 regularization, also known as least abso- operators have been combined with the idea lute shrinkage and selection operator, LASSO) behind SINDy [Brunton2016b, Klus2018]. Again, have been proposed. One prominent example this leads to methods for the identification of is sparse identification of nonlinear dynamics dynamical laws via solutions to least-squares (SINDy), aiming expressly on learning dynam- problems and allows generalizations toward ical laws [Brunton2016a]. control problems [Brunton2016b]. Recently, this kind of approach has been utilized for sparse Recent research has demonstrated that the identification by means of replacing the transfer idea of identifying dynamical laws via sparse operator by the underlying infinitesimal genera- solutions to regression and least-squares-like tor [Klus2020]. This technique, termed gEDMD, problems is impressively powerful in applica- in particular shows remarkably nice properties tions. Since 2016, several advanced variants in learning sparse representations of stochastic of this idea have been introduced (e.g., exten- dynamics (SDEs) such as agent-based models, sions to high-dimensional systems like MANDy see Figure 5. 100 250 Prey 120 Predator
100 80 200
80 ts
60 150 predators
f 60 agen ro f be
40 ro 100 40 Num be 20
20 Num 50 50 100 150 200 Numberof prey 0 0 0 20 40 60 80 100 0 200 400 600 800 1000 Time t 120
100
80 predators 60 of er 40 Nu mb 20
50 100 150 200 Numberofprey 33 Figure 5: Learning the collective dynamics of predator-prey systems. Left panel: Snapshot of the spatial state of the predator-prey system with 250 individuals at time t = 250. Predators (red dots) and prey (green) move randomly in space, if prey is within some radius (light red) of the predator, it is eaten with a certain probability and the predator reproduces; surviving prey always reproduce. Middle panel: Evolution of the collective numbers of predators and prey in the system with time (sample trajectory). Time series of such sample trajectories are used as data for learning the laws of the collective dynamics via gEDMD. Right panel: Phase diagrams of the learned dynamics (below) and phase diagram coming from a sample trajectory. Data from Jan-Hendrik Niemann, Stefan Klus, and Christof Schütte (2020) Data-driven model reduction of agent-based systems using the Koopman generator, submitted to PLOS one. All of these algorithms (see Figure 6 for a schematic inant trend in the literature was to apply off-the-shelf representation) require an efficient and robust solver optimization algorithms like LASSO. In many cases, of the underlying sparse optimization problem, most- the inherent limitations prevented the successful ap- ly one in a basis pursuit form. Until recently, the dom- plication to realistic data.
nonlinear, nonlinear, linear non-sparse sparse Understanding and Modeling Complex Biological Systems
Extended Dynamic Dynamic Mode Sparse Identification of Mode Decomposition Nonlinear Dynamics Markovian Decomposition (DMD) (SINDy) (EDMD)
Generator-based Extended Dynamic Mode Decomposition (gEDMD)
Sparse Identification of Autoregressive models Non- Nonlinear Autoregressive and Hankel-DMD Markovian Models (SINAR)
Figure 6: Methods for learning dynamical laws from data. NOISY DATA
A challenge to be faced by SINDy-like or ED- basic building block, as opposed to projections onto MD-based algorithms is inevitable noise in real-world the admissible set, which are in general much harder data. Many available algorithms produce much less to solve. In the identification of nonlinear dynamical sparse results in the presence of noise – up to com- laws, this is the case in particular if additional equal- pletely dense solutions. While this does not necessar- ity constraints such as mass preservation need to be ily affect prediction accuracy, the reliability of insight included. in physical laws extracted from these solutions suffers tremendously. Thus, we set out to develop an algo- Such CG methods have been applied to the basis rithmic approach that is robust with respect to noise. pursuit formulation of learning nonlinear dynamical 35 laws, leading to a method dubbed CINDy (condi- Conditional gradient (CG) methods are first-order se- tional gradients-based identification of nonlinear quential linear programming algorithms for convex dynamics) as an homage to SINDy [Carderera2021]. optimization problems, with the compelling feature For more details, see the feature article “Simple Ex- of only requiring the solution of linear programs as a planations, Sparsity, and Conditional Gradients.” WIND UNDER YOUR WINGS Martin WeiserMartin | [email protected] | +49 30 841 85 170
The current system of routing flights on predefined airway networks does not harvest the benefits of exploiting tailwinds in an optimal way. Dropping the restriction to airways allows fuel to be saved and
reduces CO2 emissions – an important driver in view of growing numbers of air traffic.
Established graph-based routing algorithms face con- siderable computational challenges in this setting. At ZIB, we develop and analyze hybrid approaches combining discrete optimization algorithms with op- timal control methods in order to compute globally optimal free-flight trajectories to high accuracy in an efficient way. WIND UNDER YOUR WINGS Discrete-Continuous Free Flight Planning 37 The Future Sky
At the dawn of aviation, the sky was free, but At a long-term exponential growth of 4.4% per navigation a problem. Apart from using a com- year [Oxley, Jain 2015], congestion of the air- pass, the sun, moon, and stars, only landmarks way network was unavoidable, and started to and various sorts of beacons, vision and later become more and more of a problem for at radio provided orientation. Routes were set up least a decade. On the other hand, aircraft between them, soon stacked into vertical flight navigation equipment became more and more levels, and monitored by air traffic control. In sophisticated, reducing the need for centralized this way, today’s airway network evolved – a guidance. virtual 3D graph of 100,000 “waypoints” and 600,000–700,000 “segments” on each of 20+ Can one simply abandon the airway network flight levels, on which all air traffic takes place to fly free like a bird, with a tailwind propelling (see Figure 1 for an example). A notable ex- you forward? This would eliminate congestion
Wind under Your Wings ception was the oceans, where free flight was and detours, and save time, fuel, and emissions always possible, and very well exploited to cut – a five-fold-win situation. It therefore doesn’t short or to take advantage of favorable wind come as a surprise that the EU’s Single Euro- conditions. pean Sky program has exactly this vision, and there are similar initiatives all over the world.
Figure 1: Danish airway network. Figure 2: Restriction to the graph leads to suboptimal routes (blue) compared to the continuous optimum (green).
How much can one gain? Practical experience, To harvest these benefits, flight-planning tools recent studies [Wells et al. 2021], and our own are needed that can find the quickest, most numerical examples indicate potential savings fuel efficient, etc. free-flight route taking into- ac
in fuel and CO2 emission of about 2%, and up to count the weather conditions – and a number 10% in extreme cases. As an example, consider of other side constraints such as traffic rules – in Figure 2: the optimal free-flight route (in green) reasonable time. is not only faster, but also smoother in opera- tion, as the comparison of the heading angles
of both trajectories in Figure 3 shows. 39
Figure 3: The heading angle shows that the discrete route tends to “zigzag” around the continuous optimum. Discrete Flight Planning
The current state of the art in flight planning is marked by in Figure 4. A striking example that stresses the importance discrete dynamic programming algorithms that are based of the wind is shown in Figure 5. The search space is subdi- on superfast shortest-path algorithms. These methods ex- vided into two routes, a northern direct route, and a longer ploit the characteristics of the problem to derive powerful southern corridor, which can take advantage of strong jet- lower bounds that effectively prune the search space, such stream winds, and, despite a detour of 1,250 km, turns out that globally optimal solutions can be computed to any to save 45 minutes of flight time, 5.5 tons of fuel, 18.8 tons
desired degree of accuracy, with respect to several objec- of CO2 and $1,100 of overflight fees. tives, and subject to complex side constraints, including route availability, terrain clearance, alternate airport rules, For some time now, gradually expanding free-flight op- etc. For example, bounds from so-called super-optimal portunities have been successfully approximated by graph winds, developed at ZIB in collaboration with Lufthansa densification (see Figure 6). However, this approach does Systems [Blanco et al., 2016] can be used to significant- not scale well, such that, at some point, the problem be-
Wind under Your Wings ly tighten traditional A* bounds, which in turn already do comes intractable. much better than a pure Dijkstra approach, as illustrated
Figure 4: Search spaces of Dijkstra’s algorithm (gray) and A* (yellow). Figure 5: The search space is divided following two promising strategies. Direct connection (top) or a detour while making use of the strong jet stream (bottom).
Indeed, the run time of discrete shortest-path al- of edges: , such that the overall run- gorithms depends on the density of the underlying time becomes a mere , even without |E|=O(|V|) network. On sparse graphs, shortest paths can preprocessing. Naive free-flight networks, however, 41 O(|V| log|V|) be computed extremely fast by one of several do- have a considerably larger number of vertices and main-specific superfast algorithms. These all have a edges (see again Figure 6). This is an worst-case time complexity of , enormous 2difference, which at some point turns the |E|=O(|V| ) where and are the number of edges and computation of globally optimal shortest paths at O(|E| + |V| log|V|) vertices, respectively. The airway network, like a high resolution and accuracy into an impossibility. |E| |V| planar graph, even has a favorably low number
Figure 6: Free-flight approximation of the Danish airspace by using a denser graph. Continuous-Flight Planning
Continuous methods of optimal control, which have
been studied for a century, are an algorithmic alter- 0.4 native. Instead of covering the airspace by a graph,
they only discretize the trajectory itself, exploiting nec- 0.2 essary optimality conditions to move the waypoints continuously in space (see Figure 7). Established 0.0 methods based on Pontryagin’s maximum principle or direct collocation approximate the trajectories by – 0.2 higher-order polynomials, and thus translate the op- timality conditions into large, sparse, and structured equation systems to be solved by Newton’s method. – 0.4 This allows much coarser discretizations to achieve
Wind under Your Wings the same accuracy as graph-based approaches, – 0.6
and, combined with fast quadratic convergence, 0.0 0.2 0.4 0.6 0.8 1.0 highly accurate free-flight trajectories to be computed in a much shorter time frame. Figure 7: Newton iterates of a discretized trajectory converging quadratically toward the optimal solution.
The downside of that approach is that these methods 0.4 only converge to some nearby local optimum, if at all, and thus provide no guarantee of global opti-
0.2 mality (see Figure 8). Thus, in itself, optimal control approaches are insufficient for free-flight planning.
0.0
– 0.2
– 0.4
– 0.6
0.0 0.2 0.4 0.6 0.8 1.0
Figure 8: Empirical domain of convergence to the globally optimal trajectory. Starting Newton’s meth- od from an initial trajectory within the green region, it converges to the corresponding optimal solution. Figure 9: The optimal trajectory from Brunei to Jeddah exploits tail- wind (green) and avoids headwind regions (red) as far as possible. 43
Can discrete and continuous methods be merged into an approach that combines their strengths and eliminates their weaknesses, resulting in a method that provides, at the same time, global optimality and rapid convergence to an optimal free-flight trajectory at any desired accuracy? Exactly this is the aim of a new method that is currently developed at ZIB: we recently proposed a novel hybrid algorithm called DisCOptER [Borndörfer/Danecker/Weiser2021]. A Hybrid Algorithm
While discrete, graph-based approaches effi- 0.4 ciently approximate global optima at low to medium accuracy, continuous optimal control 0.2 methods excel at fast local convergence to high accuracy. The hybrid DisCOptER method 0.0 exploits these properties in a two-stage ap-
proach. In a nutshell, it starts an optimal control – 0.2 algorithm from a suitable path provided by a
discrete dynamic programming method. – 0.4
0.0 0.2 0.4 0.6 0.8 1.0 First, an artificial, fully connected graph is creat- ed (cf. Figure 10, blue dots) that contains origin 0.4
Wind under Your Wings and destination vertices. Then, a shortest path on this graph is calculated (red path). 0.2
Finally, this shortest path is used as an initial 0.0 guess for the following refinement stage of the
algorithm. Here, an optimal control approach – 0.2 involving some variant of Newton’s method is
used to calculate a highly accurate solution – 0.4
(green trajectory). Provided that the first guess 0.0 0.2 0.4 0.6 0.8 1.0 is within the radius of convergence of the actu- al continuous optimum (green area), Newton’s Figure 10: Continuous solution found if initial- method converges extremely fast. ized with different discrete routes. Figure 11: Time complexity of the purely discrete approach vs. the hybrid algorithm . 6 (O(1/l )) (O(1/l)) 45
For the second stage to converge reliably It turns out that in practically relevant exam- to the global optimum in the vicinity of the ples, the required graph density is rather path provided by the first stage, this path low (again, see Figure 10). Consequently, must be sufficiently close to the global opti- the discrete shortest path can be found mum, more precisely, within the domain of quickly, the problematic asymptotic runtime convergence of Newton’s method. This can behavior of the discrete graph search is cir- be ensured by making the first-stage graph cumvented, and the overall time complexity sufficiently dense. The required graph reso- is governed by the asymptotically efficient lution depends only on external factors like optimal control stage (cf. Figure 11, is a the wind conditions, but not on the desired measure for the solution accuracy). But can 1/l accuracy, which only affects the trajectory we decide within the algorithm how dense discretization in the second stage. the graph needs to be in a concrete case? Error Estimates for Discrete Paths
For free-flight trajectory planning, the graph used either Following the standard paradigm of a priori error esti- stand-alone for path planning or as the first stage in a mates for finite element discretizations [Deuflhard/Weiser, hybrid algorithm needs to be sufficiently fine for the dis- 2020], the distance of an optimal continuous path to its crete optimal path to be accurate enough or to lie within interpolate in the graph is bounded, in terms of the flight the convergence domain of Newton’s method, respectively. duration, by bounding the second derivative of the objec- To ensure this, a priori error estimates for optimal flight tive with respect to path deviations. The structure of - paths in locally densely connected graphs have been dense graphs guarantees the existence of a reasonable (h,l) derived [Borndörfer/Danecker/Weiser2021-b]. -dense interpolant. Since the optimal discrete path in the graph graphs are those where the h-balls around vertices cover is not worse than the considered interpolant, this provides (h,l)
Wind under Your Wings the plane, and all vertices with a distance of at most an error bound for the discrete solution in terms of vertex are connected by an edge (see Figure 12). density , local connectivity length , and magnitudes of l+2h wind speed and its spatial derivatives. h l
The excess in flight duration is due to spatial displace- ment of the trajectory on one hand, and zigzagging due to limited angular resolution on the other hand. The latter is bounded by , whereas the former is controlled by (see Figure 13). Based on this distinction, the error estimate l/h h allows the selection of a theoretically optimal relation of vertex density and connectivity radius , which turns out to be . The resulting error bound guarantees a h l convergence 2of optimal discrete paths toward the contin- h = O(l ) uous optimum in terms of flight duration at a rate of . 2 O(l )
Figure 12: -dense graph.
(h,l) hybrid algorithms. curacy andallows implementing efficientand reliable guidethe graphactually refinement to the desired ac teriori error estimators can be designed, which can 14). This provides the basis onwhich efficientapos convergenceobserved order accurately (seeFigure from sharp. Nevertheless, they capture the numerically usual,As the apriori error boundsare reliable butfar
T( G) – T( C) 10 10 10 –3 –2 –1 Figure 14:Numerical experiments confirm the order of the theoretical error bound. 6 × 10 –2 10 –1 Connectivity - - a Figure 13:Angular error discrete (red) and general enough to beapplicable to many other prob principle isnot limited to path problems, shortest butis with fasttimality local convergence. The underlying crete-continuous algorithms that combine global op DisCOptER isafirst exampletype ofdis ofanew lems ofsimilarnature.
2 × 10 a continuous route (green). –1 (l) anddistance error 3 × 10 –1 4 (r) × between between 10 –1 - - -
47 NHR@ZIB: TAKING THE NEXT STEP IN HPC
The NHR Center at ZIB provides high-performance compute and storage resources to support the German science community. Thomas Steinke | [email protected]
SGI ICE, XE Atos/Intel | +49 30 841 85-144 Cray T3E 10,240 cores 121,920 cores Cray X-MP 256 cores Intel Harpertown, Intel Cascade The German federal and 1984 1994 2002 2013 state governments decid- 2 cores DEC Alpha Nehalem Lake AP ed to jointly fund the Na- 471 MFlops 486 GFlops 160 TFlops 8.2 PFlops tional High Performance Computing Alliance (NHR). This will create a national- ly coordinated network of Berlin‘s High-Performance Computers at ZIB HLRN Supercomputers at ZIB high-performance comput- ing centers. Eight centers, including ZIB, were select- ed on the basis of a com- petitive and science-driven Cray 1M Cray T3D IBM p690 Cray XC40/30 process involving a review 1 core 192 cores 384 cores 44,928 cores by the German Research 160 MFlops 1987 DEC ALpha 1997 IBM Power4 2008 Intel Ivy Bridge, 2019 Foundation (DFG) and an evaluation by an indepen- 38 GFlops 2.5 TFlops Haswell dent strategy committee 1.3 PFlops appointed by Germany’s Joint Science Conference. For the next ten years, funding of more than 70 million euros was awarded for the NHR Center at ZIB. NHR@ZIB: TAKING THE NEXT STEP IN HPC
SGI ICE, XE Atos/Intel 49 Cray T3E 10,240 cores 121,920 cores Cray X-MP 1994 256 cores 2002 Intel Harpertown, 2013 Intel Cascade 1984 2 cores DEC Alpha Nehalem Lake AP 471 MFlops 486 GFlops 160 TFlops 8.2 PFlops
Berlin‘s High-Performance Computers at ZIB HLRN Supercomputers at ZIB
Cray 1M Cray T3D IBM p690 Cray XC40/30 1 core 192 cores 384 cores 44,928 cores 160 MFlops 1987 DEC ALpha 1997 IBM Power4 2008 Intel Ivy Bridge, 2019 38 GFlops 2.5 TFlops Haswell 1.3 PFlops HPC Services at ZIB – a Long Journey
Scientific computing, nowadays complemented by in the context of NHR. So, since the first days, an artificial-intelligence applications, is the undisputed essential aspect has been the seamless integration third pillar of science today. This is the result of evolu- of scientific and technical expertise with the excellent tion over the last decades in many areas of science, administrative support of users across different sci- technology, manufacturing, and, last but not least, the ence domains. Physics, chemistry, and engineering enthusiasm of young scientists and engineers in rec- evolved as traditional HPC science domains in the
NHR@ZIB ognizing the potential of performing experiments “in 80s and therefore the early team of HPC consultants silico.” In the beginning, high-performance computer at ZIB included senior scientists from these areas. or supercomputer systems were special designs and Nowadays, other science disciplines like life science used by a few groups of scientists who needed to and data science have joined. understand the hardware and software technology behind the scene. Operation of supercomputers was Sharing different supercomputer models and nontrivial from the beginning, and the need to pro- best-practice solutions among different academic vide special support to the domain scientists soon HPC sites is another promising way to cope with the became clear to ensure that these powerful but ex- increasing demand for HPC services. As early as with pensive resources were used efficiently. its newly installed CRAY 1M in 1984, ZIB together with the universities in Kiel and Hannover founded Consequently, in 1984, the Zuse Institute Berlin was the North-German Vector computer Network (NVV)1 founded as an institute for scientific computing as to establish one of the first HPC alliances across Ger- well as a provider of high-performance computing man states. The NVV not only shared the hardware (HPC) resources. From the very beginning, an impor resources, at that time dominated by vector-computer tant goal was to establish first-class scientific sup- architectures with new installations rotating every two port provided by HPC consultants to promote HPC years, but also a distributed network of HPC consul- in the science communities in Berlin; later, this was tants was implemented together with cross-site user expanded to northern Germany, and now nationwide management of the joint-user base.
1 The NVV (German: Norddeutscher Vektorrechnerverbund) was based on a state treaty between the federal states of Berlin, Schleswig-Holstein, and Lower Saxony. In 2001, the next major step was taken: The transregional and interdisciplinary competence North German Supercomputing Alliance2 network consisting of application experts. This (HLRN) was founded. With the experiences of network involved all boards and institutions in the NVV having been well accepted by scien- the HLRN. tists and uniquely visible by German federal and state bodies, the seven federal states So far, four generations of HPC systems have Berlin, Brandenburg (since 2012), Bremen, been operated, with differences in their com- Hamburg, Mecklenburg-Western Pomerania, puter architecture including processor models
Lower Saxony, and Schleswig-Holstein now and high-speed network types (see box). The 51 became partners to foster HPC for researchers “Lise” system is the current flagship system in northern Germany. The HLRN alliance jointly at ZIB with a peak performance of about operated a distributed supercomputer system 8,000,000,000,000,000 (Peta) floating-point op- at the two HLRN sites: one HPC system at ZIB erations per second (PFLOPS). Such a class of in Berlin and its sister system in Lower Saxony HPC system provides many levels of concurren- at the Leibniz University Hannover (until 2018) cy of processing data to achieve this tremen- and the Georg-August-Universität Göttingen dous performance. And clearly, it is not easy to (from 2018), respectively. With the joint forces exploit this potential as it requires a sufficient of seven states, not only very competitive HPC understanding of the system architecture, the system resources could be provided to the software tools, and application packages. scientific-research community in Germany, but Again, this is part of the daily scientific services the users of the HLRN could also resort to a of the HPC team at ZIB to the users.
2 Norddeutscher Verbund für Hoch- und Höchstleistungsrechnen – HLRN. The Next Step: The NHR Center at ZIB
Over the last half decade, it has become evident that In order to meet the increasing importance of and demand nationwide further steps have to be taken to foster and for high-performance computing, in November 2018, the extend high-performance computing infrastructure and German Joint Science Conference3 (GWK) agreed on the competence at a broader level in the German science joint funding of a coordinated National High-Performance community. Not only the increasing demands for compu- Computing Alliance (NHR) by the federal and state govern- tational resources has become increasingly challenging, ments. Importantly, in addition to the provision of comput- but also the skill set now required for using these resourc- ing capacities, one focus of the NHR network, according to es efficiently has required more and more attention. For the agreement, was the strengthening of methodological example, as laptops and smartphones become parallel competence through coordinated education and training computers at the lower end, and large, highly parallel, and of users and, in particular, young scientists. very powerful computer systems at the top develop toward
NHR@ZIB exascale computing, developing high-quality scientific In January 2020, the call for proposals to become a software for these machines becomes increasingly more NHR Centers was published. In an open, competitive, complex, as a deep technical understanding is a must. The science-driven process involving a review by the German gap between the performance capabilities of the hard- Research Foundation (DFG) and an evaluation by an inde- ware and the performance achieved with real-world codes pendent strategy committee appointed by the GWK, the has to be minimized. Simultaneously, the gap between the new Tier-2 HPC centers were selected. In November 2020, required skills for using Tier-3 (universities, research groups) the GWK announced eight HPC centers as new NHR mem- and Tier-1 (pre-/exascale) systems has to be closed, too ber sites with the ZIB as one of them. The NHR Center at – a call for better coordinated actions for tier-2 HPC sites ZIB has been in operation since January 2021. is needed. Furthermore, the operational costs of a Tier-2 HPC system over its five- to six-year life span become as expensive as the amount invested.
3 Gemeinsame Wissenschaftskonferenz – GWK. The Next Years as NHR Center
The HPC world is facing a few challenges in the coming tivity steps of the NHR centers, ZIB, the Friedrich-Alexander years. The NHR network offers the potential for solving University Erlangen-Nuremberg (FAU), and the Paderborn some of them by combining the diverse expertise and University (UPB) are establishing the NHR Competence working jointly on some of the demanding challenges. At Center for Atomistic Simulations to support the broad the technological and methodological level, one important scope of atom-based simulation techniques in the areas of challenge is how future Tier-2-level HPC system architec- physics, chemistry, and engineering. As a horizontal meth- 53 tures will look like. The answer to this is mostly driven by od, machine learning/artificial-intelligence workflows will economical demands to operate HPC facilities. Heteroge- be increasingly integrated into the model-driven simulation neity is one promising key word that is part of the solu- schemes. For the next-generation HPC architecture, one tion. Diversity in the worlds of processors and accelerators can try to maximize the sustained application performance (general-purpose graphic processing units, vector engines, for important software packages of the science domains field-programmable gate arrays, special chips for artificial highlighted above. Such a heterogeneous system architec- intelligence operations, quantum computers) and memo- ture may significantly comprise GPUs that are complement- ry technologies (e.g., high-bandwidth memory, traditional ed by FPGAs or available AI processors. So, as the next DRAM, storage-class memory, memory over fabric) offer challenge, the question arises as to what extent existing greater space for designing innovative system architectures. application software can take advantage of such a het- But there is no one-size-fits-all architecture providing maxi- erogeneous system architecture, and which program codes mum performance while staying in a given power envelope. that are important in the science communities need to be NHR offers the potential for a coordinated and balanced migrated and optimized for these next-generation systems. specialization in terms of future system architectures and the Again, the NHR Alliance provides opportunities for collab- way the science domains, with their software codes, can be orative work on optimizing such codes here, for example, supported on the NHR systems. for accelerators. With FAU, UPB, and RWTH Aachen, ZIB is initiating an open NHR project called “Performance Lab” The NHR Center at ZIB will take the lead in selected in which these problems can be tackled. application domains (e.g., life sciences with a focus on model-driven simulations, advanced integration of machine Funding for the NHR Alliance was secured until 2030. For learning and simulation, or chemistry/material science) to- the next ten years, we are looking forward to welcoming gether with other NHR centers. As one of the first joint-ac- the NHR user community at ZIB. SIMPLE EXPLANATIONS, SPARSITY, AND CONDITIONAL GRADIENTS
Christoph Spiegel | Motivation
Almost all questions can have a complex answer or a simple answer. This rings true in particular in science, and we typically strive to find the most
[email protected] parsimonious explanation, i.e., the simplest explanation that is consistent with an observed phenomenon: | +49 30 841 85 - 436 “EVERYTHING SHOULD BE MADE AS SIMPLE AS POSSIBLE, BUT NO SIMPLER.”
Albert Einstein. 55 Simple Explanations, Sparsity, and Conditional Gradients A simple explanation is more stable, reacts better to very non-convex deep neural networks. For many ap- noise, and does not induce “structure” or “phenome- plications, the parameter space is often constrained, na” beyond what is in the data. Moreover, also from either due to natural restrictions or because a partic- a Bayesian statistics perspective, the most parsimoni- ular regularizing effect is desired. Frank-Wolfe algo- ous model is (almost) always the “best” answer. Now rithms (Frank and Wolfe, 1956), one of the simplest that we are entering an age where we use more and earliest known classes of iterative algorithms and more machine learning (ML) systems to gain in- and also often referred to as conditional gradient sights into the world, this imperative is as important algorithms (Levitin and Polyak, 1966), have been the as ever: nobody can interpret the more than 1 billion center of enormous renewed interest over the last parameters of modern deep-learning systems. While several years (see, e.g., Jaggi, 2013) and have in- we do often need a large number of parameters to creasingly established themselves as the method of describe a phenomenon, we would still like to ad- choice for such constrained optimization problems here to the principle of the simplest sufficient model when we seek sparsity of the final solution. Not only in order to explain what is going on. can they handle constraints very efficiently as they maintain feasibility of the iterates without requiring At the core of ML training is often the solution to an projections, they also naturally induce sparsity of the optimization problem and the parsimoniousness from final solution as it is built up as a convex combina- above translates into the “sparsity” of the solution. tion of a few atoms. Additionally, these methods are For example, the famous LASSO regression problem computationally very efficient, so that rather large aims for regression, where the number of explain- learning problems can be solved efficiently and ing factors in the final solution is kept small. More in particular the computational cost is often much generally, one might strive for solutions to these ML cheaper than that of corresponding projection-based optimization problems that are made up of few (po- methods (Combettes and Pokutta, 2021). tentially complex) base objects; these base objects are often called atoms. First-order methods are es- In the following, we present a few examples of our sential to many modern optimization problems, from work in the area of conditional gradients with a par- smooth and convex regression problems in lower-di- ticular focus on sparsity. mensional spaces to highly overparameterized and Conditional Gradient Algorithms for Learning Dynamical Systems
As mentioned earlier, the modern age of machine Conditional gradient methods are an exceptionally learning and big data that we live in has enabled attractive family of algorithms for these types of appli- the rise of deep-learning models, often possessing cations because of their sparsity-inducing properties. billions of parameters, which can be trained to clas- They produce more accurate and sparser solutions sify or predict with high accuracy when given enough than other competing algorithms in the presence input data. However, these models are oblivious to of noise, allowing us to find governing differential how the underlying data with which they are trained equations with less training data, for example. This is generated, and often fail to predict or classify out- means that the learned governing equations often of-sample data, coming from a regime that has not contain fewer spurious terms and noise is less likely 57 been encountered in the training process. to be explained by phantom terms. Moreover, due to the fact that these families of methods are easily This approach stands in contrast to how many of the able to deal with linear constraints (and more gen- phenomena in physics were described or explained erally convex constraints), we can add constraints to in previous centuries, that is, through the use of dif- our learning problem to impose certain symmetries ferential equations. These equations have helped on our learned model, or to guarantee certain con- build our understanding in fields as wide-ranging as servation properties are satisfied, thereby producing quantum mechanics, fluid mechanics, or electromag- models that are consistent with the underlying physi- netism, and have enabled a deep understanding of cal phenomena (Carderera et al. 2021). We call this how the data coming from a physical system has conditional gradient-based recovery method CINDy been generated. Given the current availability of (short for conditional gradient-based identification of data, there has been a recent trend to find these dif- nonlinear dynamics), in homage to SINDy (Brunton et ferential equations through sparse regression meth- al., 2016). On numerical examples with noisy data, odologies. Such methods attempt to find the best CINDy outperforms several alternative approaches sparse fit to observational data in a large library in terms of (1) prediction accuracy, (2) sparsity, and of potential nonlinear models and often perform ex- (3) replication of dynamics trajectories. One reason tremely well with noise-free data but fail to produce is that CINDy blends orthogonal matching-pursuit-like accurate or sparse solutions when noise is present. steps gradually into basis-pursuit-like steps by picking Please see the feature article “Understanding and up atoms sequentially, which are only accepted into Modeling Complex Biological Systems” for more de- the set of active atoms if the improvement of training tails on such methods for learning dynamical laws accuracy significantly outweighs the loss in sparsity. from data. This leads to extremely sparse iterates, often close to the information-theoretic recovery limit. The Kuramoto system, consisting 2,000 of a set of weakly coupled oscilla- tors with slightly different natural 1,000 frequencies, is a model for study- ing synchronization phenomena. –7 –5 –3 The natural dictionary contains 10 10 10 trigonometric functions and is Noise level overcomplete. This results in other Figure 1: Number of wrongly approaches selecting too many selected atoms for the Kuramo- Simple Explanations, Sparsity, and Conditional Gradients to example over the noise level atoms and producing dense solu- in the trajectory data. tions (Figure 1). The extra atoms do not only hide the actual struc- ture of the Kuramoto model, but also lead to inferior model predic- tions and evolution of trajectories (Figures 1 and 2).
Figure 2: Trajectory snapshots of dynamics learned by the SINDy and CINDy algorithms with data from the Kuramoto model under the presence of noise. The ordinary differential equation learned by the CINDy algorithm produces a trajectory that well matches that of the exact dynamic. Conditional Gradients for Everyone
As the aforementioned conditional gradient algorithms 3. It should work at scale. In order to work are an essential part of our own work but are also at scale for very large-scale models with heavily used within the broader research community, several billion variables, memory manage- we are currently developing a new high-performance ment becomes very important. At that scale, software library implementing major Frank-Wolfe vari- allocations of memory cost real time and ants in Julia. The algorithms can be used from within you are close to the physical memory limit other common programming languages (such as C, of modern machines, so that unnecessary C++, Python) and frameworks (such as MathOptInter- allocations can easily lead to out-of-memory face and JuMP in Julia) through respective wrappers. blowups. We paid particular attention to this by supporting a memory-emphasis mode in We had several specific design aspects in mind, when all algorithms that significantly reduces the designing the toolbox. In particular, we tried to push memory consumption. We also implemented the envelope on the usual “speed vs. robustness vs. a variety of constant-storage sparse repre- scale” trade-off triangle: sentations for common feasible regions (in- dependent of the problem dimension). 1. It should be fast. That is why we opted for Julia in first place asthe language and the 4. It should be flexible. We support almost
overall design is geared toward speed with arbitrary number types and element types 59 specialized data structures. for the iterates out of the box. Notable ex- amples are rational arithmetic, so that we 2. It should be robust. When solving optimi- can, for example, solve the approximate Car- zation problems, numerical instabilities can athéodory problem to actual optimality. be a huge problem when high-accuracy solutions are sought after, or the data is 5. It should be easy to use. The best code ill-conditioned. We cater to this by support- is worth nothing if it cannot be used. As ing various floating-point modes out of the such, we provide ample interfaces and hide box allowing for very-high-precision com- all complexities within the framework, so that puting (e.g., using BigFloats). the user only needs to supply a linear op- timization oracle for the feasible region as well as a first-order oracle for computing gra- dients. Gradient computations can also be transparently performed using GPUs or other hardware accelerators. To put things into numbers, the following table provides statistics for various components of the basic Frank-Wolfe algorithms, here solving a LAS- SO problem over 1e9 (i.e., one billion variables).
Size of single vector (Float64): 7629.39453125 MB
Time Allocations
Tot / % measured: 278s / 31.4% 969GiB / 30.8% Simple Explanations, Sparsity, and Conditional Gradients Section ncalls time %tot avg alloc %tot avg
update (blas) 10 36.1 s 41.3% 3.61 s 149 GiB 50.0% 14.9 GiB
lmo 10 18.4 s 21.1% 1.84 s 0.00 B 0.00% 0.00 B
grad 10 12.8 s 14.6% 1.28 s 74.5 GiB 25.0% 7.45 GiB
f 10 12.7 s 14.5% 1.27 s 74.5 GiB 25.0% 7.45 GiB
update 10 5.00 s 5.72% 500 ms 0.00 B 0.00% 0.00 B (memory)
dual gap 10 2.40 s 2.75% 240 ms 0.00 B 0.00% 0.00 B
Each gradient here is about 7.6 GB of memory and a single gradient computation takes about 1.28 seconds. We can also see that while an update of the iterate in optimized linear al- gebra requires about 15.2 GB of memory, in memory emphasis mode, it can be computed without allocating memory, so that it ends up being faster than the optimized linear algebra operations at that scale due to the large cost of allocating memory. Vanilla Frank-Wolfe Algorithm.
EMPHASIS: Memory STEPSIZE: Agnostic EPSILON: 1.0e-7 MAXITERATION: 1,000 TYPE: Float64 MOMENTUM: Nothing GRADIENTTYPE: Nothing WARNING: In memory emphasis mode iterates are written back into x0!
Type Iteration Primal Dual Dual Gap Time It/sec
I 0 1.000000e+00 –1.000000e+00 2.000000e+00 8.456415e+00 0.000000e+00
FW 100 1.326732e-02 –1.326733e-02 2.653465e-02 4.628479e+02 2.160537e-01
FW 200 6.650080e-03 –6.650086e-03 1.330017e-02 9.171626e+02 2.180638e-01
FW 300 4.437059e-03 –4.437064e-03 8.874123e-03 1.371521e+03 2.187352e-01
FW 400 3.329174e-03 –3.329180e-03 6.658354e-03 1.825769e+03 2.190858e-01
FW 500 2.664003e-03 –2.664008e-03 5.328011e-03 2.280048e+03 2.192936e-01 61
FW 600 2.220371e-03 –2.220376e-03 4.440747e-03 2.734262e+03 2.194376e-01
FW 700 1.903401e-03 –1.903406e-03 3.806807e-03 3.188507e+03 2.195385e-01
FW 800 1.665624e-03 –1.665629e-03 3.331253e-03 3.642893e+03 2.196057e-01
FW 900 1.480657e-03 –1.480662e-03 2.961319e-03 4.097155e+03 2.196646e-01
FW 1,000 1.332665e-03 –1.332670e-03 2.665335e-03 4.551414e+03 2.197119e-01
Last 1,000 1.331334e-03 –1.331339e-03 2.662673e-03 4.556530e+03 2.196847e-01
4557.523048 seconds (6.90 M allocations: 112.096 GiB, 0.01% gc time)
Here, these 1,000 iterations were performed in is that we aggressively reuse memory, so that the about 4,557 seconds (i.e., about 4.55 seconds per only real memory allocations that are required iteration requiring about 112.096 GB of memory are those for computing information for reporting. allocation, which works out to just 112 MB per iter- It should also be noted that conditional gradi- ation), which is considerably less than the size of ents offer dimension-independent convergence a single vector (7.629 GB) and we maintain two guarantees, with consistent convergence speed of these: the gradient and the iterate. The reason irrespective of the problem size. Conditional Gradients in Deep Learning Simple Explanations, Sparsity, and Conditional Gradients
Another natural application of condition- 1 al gradient methods is in the context of SGD L -norm ball deep learning, where the sparsity-induc- ing properties of these methods – togeth- er with an appropriately chosen feasible region – can significantly affect how in- 0 formation is represented within the neural network and how decisions are made by the network. For example, in Figure 3, we can see that depending on the method and feasible region used for training, the network relies more on positive evidence (green pixels) or negative evidence (red 5 pixels). Here, evidence relates whether the presence of a pixel (green) is evi- dence for a given class (here the digits 0 or 5) or whether the presence of a pixel (red) is evidence against a given class. Figure 3: Visualization of the weights in a fully connected As such, conditional gradient algorithms no-hidden-layer classifier trained on the MNIST dataset can significantly change the statistical corresponding to the digits 0 and 5. Red corresponds to negative and green to positive weights. properties of the trained neural network altering, for example, the false-positive vs. false-negative trade-off. 2 In order to be able to try our methods on state- L -norm ball Hypercube of-the-art neural-network training problems, we have likewise implemented several Frank-Wolfe methods in the two most common frameworks used for the optimization of neural networks, TensorFlow and PyTorch. In our experiments, we have demonstrated that stochastic Frank-Wolfe methods can achieve state-of-the-art test accu-
racy results on several well-studied benchmark 63 datasets, namely CIFAR-10, CIFAR-100 and Im- ageNet. We furthermore show that the chosen feasible region significantly affects the encoding of information into the networks both through a simple visualization and by studying the number of active weights of networks trained on MNIST with various types of constraints, as already dis- cussed above (see also Figures 1 and 2). RESEARCH CAMPUS MODAL: SUCCESS Ralf Borndörfer | STORIES [email protected]
Funding for Research Campus | +49 30 841 85-243 MODAL Renewed
After successful completion of its first funding phase from 2014 to 2020, the Research Campus MODAL was eval- uated by an expert review panel and the grand jury of the German Federal Ministry of Research and Education (BMBF). MODAL’s proposal for a sec- ond phase was granted (2020–2025). The grant includes 10 million euros funding by the BMBF and a consid- erably larger amount by the 35 par- ticipating private companies. MOD- AL hosts more than 60 researchers from academia and industry under one roof. This article reports on the plans for the second funding period, presents success stories from the first funding period, and highlights re- search and practice training events of MODAL. 65 Smart control of energy Optical planning for Smart solutions for digital transport networks traffic networks precision medicine
MODAL is a public-private partnership project of a proven value of several million euros per year. Two the research partners Zuse Institute Berlin (ZIB) and of these success stories advanced to the 2020 finals Research Campus MODAL: Success Stories Freie Universität Berlin with more than 30 industrial of two of the most renowned awards competitions partners, funded by the German Federal Ministry of in research transfer. We will report on these success Education and Research within the “Research Cam- stories in detail below. pus – Public-Private Partnership for Innovation” fund- ing initiative. The research campus is hosted by ZIB, The different application areas considered in MODAL and conducts research and development of digital face similar challenges requiring mathematical meth- systems for the optimization of data-driven processes odological competence with a focus on Modeling, in the fields of energy, health, mobility, and commu- Simulation and Optimization (MSO), Artificial Intelli- nication. gence (AI), and High-Performance Computing (HPC). In MODAL, these methodological core competencies The developed data-driven methods make rapid are complemented by application expertise related cross-technology development possible and advance to specific data-driven economic processes in the innovation processes by digital decision-support sys- areas of energy, mobility, health, and nanotechnolo- tems. Present projects aim at the solution of previ- gy. The resulting competence matrix for the second ously unavailable planning and control problems for funding phase is illustrated in Figure 1 and constitutes complex supply and mobility networks or for digital both the strategic research program and the unique health care. selling point of the Research Campus. MODAL is strong in communicating this unique mix of research In the first funding phase, the MODAL research cam- excellence and proven application competence in pus selected mobility, sustainable energy supply, industry to the next generation of researchers. One secure communication, and personalized medicine of MODAL’s main activities in this field, the school on as the main topics of its research program. During “Combinatorial Optimization at Work” (CO@Work), the first funding period, MODAL produced a series of was organized for the sixth time in 2020, and attract- success stories in transferring basic research in Ap- ed more than 1,000 participants . We report on this plied Mathematics into highly relevant technological event in detail below. innovations with productive utilization in industry and Smart design of resonant Solutions for High-performance nanostructures high-performance optimization optimization software
Figure 1: In its second funding phase, the Research Campus MODAL comprises six labs. The EnergyLab is based on the former GasLab, the MobilityLab extends the former RailLab, while MedLab and SynLab just shifted focus. With HPCLab and NanoLab, new aspects were introduced.
SPECIFIC APPLICATION COMPETENCIES
SUSTAINABLE CONNECTED PERSONALIZED SECURE ENERGY SUPPLY MOBILITY MEDICINE COMMUNICATION 67
Modeling, Simulation, Optimization (MSO)
Data Management and Analysis
Artificial Intelligence and Machine Learning
SYNLAB High-Performance Computing
HPCLAB Efficient Implementation on NextGen Hardware
ENERGY LAB MOBILITY LAB MED LAB NANO LAB
CORE COMPETENCIES IN THE FIELD OF OPTIMIZATION OF DATA-DRIVEN PROCESSES
Figure 2: Mobility, sustainable energy supply, secure communication, and personalized medicine are the main topics of MODAL’s research and research transfer program. Despite this broad range of topics, MODAL is characterized by strict coherence in terms of content, which results from the mathematical methodological competence with a focus on Modeling, Simulation, Optimization (MSO), and AI. These core competencies (shown horizontally here) are complemented by (vertical) application competencies related to specific data-driven economic processes. The resulting com- petence matrix illustrates both the strategic research program and the unique selling proposition of the research campus. MobilityLab Enters Finals of Franz Edelman Competition 2020 – Mobility
A MODAL team with members of Deutsche Bahn (DB), ZIB, and LBW Optimization advanced to the final of the renowned Franz Edelman Award 2020 for Achievement
Research Campus MODAL: Success Stories in Operations Research and the Management Sciences with FEO, a decision-support system to schedule the train rotations of DB. Based on completely new mathemati- cal methods of hypergraph optimization and efficient algorithms based on it, both developed within MODAL, FEO has greatly increased DB’s planning agility and pro- ductivity, and produced significant direct savings of roughly 75 million euros per year.
The Franz Edelman Award for Achievement in Oper- The combined impact of all Edelman finalist contri- ations Research and the Management Sciences is butions since 1974 is impressive, amounting to a total the “Nobel Prize of Operations Research.” Present- of $302 billion. The contributions cover a range of (up ed by the Institute for Operations Research and the to today) 143 different application areas. In this way, Management Sciences (INFORMS), it is organized as the records of the Edelman Award serve not only as an international competition of successful OR appli- an objective indicator of a hard, monetary “Value of cations. A winner is chosen among six finalist teams, Operations Research,” but also as a showcase for who present their contributions in an illustrious gala the practical viability of groundbreaking new math- at the INFORMS Conference on Applied Analytics. ematical ideas. The 2020 Edelman gala was originally scheduled to cusing on using the best possible OR methods. FEO take place in Denver. Due to the coronavirus situa- is designed to deal with complex and large-scale tion, however, it was rescheduled and held, for the scenarios. DB is the largest European cargo and first time, as a virtual online event on September 29. regional passenger train operator, and the second Five teams participated in the final: Carnival & plc, largest in long-distance passenger rail transport, car- the world market leader in cruises, with the YODA rying 2 billion passengers per year. DB Cargo owns (Yield Optimization and Demand Analytics) revenue over 2,500 locomotives (shunters included), distribut- management system; IBMServices (the Global Tech- ed over 17 European countries, of which around 800 nology Services Unit within IBM) identified devices at locomotives serve the core network in central Europe. risk from outages before they occur; Intel improved DB Regio owns over 4,000 railcars and over 850 lo- 69 their supply chain; Walmart developed an innova- comotives (shunters included). DB Fernverkehr owns tive markdown approach for perishable goods; and over 270 intercity express (ICE) high-speed railcars Deutsche Bahn scheduled its 26,000 trips per day and over 240 locomotives (shunters included). At a using the FEO (Fleet Employment Optimization deci- price of up to 33 million euros per piece (ICE 3 type sion-support system. 407/Velaro), the efficient use of this rolling stock is critical. FEO is designed as a modular software system centered around adaptable optimization kernels, fo-
Figure 3: A team comprising Deutsche Bahn, LBW Optimization, and ZIB advanced to the final of the 2020 Franz Edelman Competition. Figure 4: Timetabled ICE 1 trips in a cyclic standard week (left). FEO GUI screenshot (right).
A key problem is train rotation planning (i.e., the Traditionally, train rotations are planned in an incre- construction of the rolling-stock rotations in order to mental fashion that was straightforward to imple- implement a given timetable [BK+18]). In contrast to ment. Today, as a result of the deregulation of the public transit or air traffic [BGJ10], where individual European railway sector, train services change much Research Campus MODAL: Success Stories vehicles can be scheduled using network-flow ap- more dynamically due to competition, and resource proaches, several units of rolling stock are combined allocation is mainly driven by cost minimization. This to form trains. The types (e.g., country-specific power results in highly challenging scheduling problems in or safety systems), positions (e.g., restaurant in the large networks, with complex technical constraints, middle), and orientations (e.g., tick = first class in over multiple time scales, integrating several stages front, tock = second class in front) of the units in a train of operations. These could not be tackled using es- follow elaborate rules and are not easily changed. tablished methods. In fact, the inability to solve the This means that vehicle rotations and train composi- key problem of train rotation scheduling was a major tions have to be planned simultaneously in order to obstacle that prevented OR methods for the longest arrive at an implementable solution. The same holds time from blazing a trail of success in the railway for maintenance intervals and parking assignments. industry comparable to the one in the airline industry. The main objectives are cost minimization (number of vehicles and deadhead mileage) and regularity The FEO optimization core is based on completely (i.e., similarity of operation over several days of the new mathematical methods of algorithmic hyper- planning horizon). Typical planning scenarios include graph theory developed within MODAL at ZIB [BR+11, the cyclic “standard week” in long-term planning (cf. BR+12]. This approach ensures the integrated treat- Figure 4 left) and fully dated planning over four to ment of the main operational requirements of train-ro- six weeks. tation creation, train composition, maintenance inter-
composition configuration vehicle timetable layer layer layer input
Figure 5: Idea behind the hyperflow train-rotation model (left). The three layers of the coarse-to-fine method model compositions, configurations, and vehicle flows (right). Figure 6: The ICE rotations through the 2015 Cologne-Rhine-Main construction site were managed by FEO (square left and right). The impact of FEO was analyzed for four dedicated construction sites marked D, D, O, U (left).
vals, parking, and regularity, all within one generic Fernverkehr to drive roughly one million additional hypergraph model (cf. Figure 5 left). Exploiting the train kilometers with the existing fleet. This equals the special “graph-based” structure is the key to the de- annual mileage of two ICE double-traction train sets velopment of a novel coarse-to-fine (C2F) algorithm and would generate additional fare revenue of ap- for the solution of large-scale instances with up to proximately 24 million euros per year or equate to the one-hundred-million hyperarcs (= variables) [BRS14] transport of 784,000 additional travelers. Each train
(cf. Figure 5 right). C2F works on coarse layers where journey saves around 44 kilograms of CO2 in compar- possible and resorts to finer layers when needed ison to the amount that would have been expended to compute a provably optimal linear programming had each passenger driven an automobile. Thus, FEO solution and in turn an integer solution of provable saves approximately 34,000 tons of CO2 emissions per quality [BR+15]. The hypergraph train-rotation mod- year. 71 el also gave rise to deep theoretical developments in combinatorial optimization, including a general- DB Regio uses FEO in managing public tenders for ized Hall theorem for normal hypergraphs [BB18], a regional rail services all over Germany. In these ten- tight-cut-decomposition algorithm for matching cov- ders, the train-operating company that requires the ered uniformizable hypergraphs [B19], and a purely lowest subsidy wins the contract for a period of up combinatorial hypergraph network simplex algorithm to 15 years. Three cases were documented in which [BG+17, BG+17, B18]. Since 2014, the transfer re- FEO was instrumental in winning tenders in the state search within MODAL’s RailLab has been focused on of Baden-Württemberg (see Figure 7), which together developing efficient algorithms based on hypergraph cover mileage of roughly 11 million train kilometers theory; these algorithms are now in productive use per year and an annual revenue of 23 million eu- in the core of FEO. ros. In all cases, pivotal optimization opportunities were identified by human ingenuity, and their precise DB Fernverkehr first used FEO for supporting decisions impact analyzed using FEO: in one case, the reposi- about the fleet and train sizes of the new ICE4 fleet. tioning of a vehicle could be squeezed into a reserve Since the rollout of FEO, its main use has been in the time, and in another, using vehicles with superior ac- management of construction sites, currently 850 per celeration capabilities gave room to implement an day. The impact of FEO was analyzed in four ded- operational concept involving heavy coupling and icated cases from the fourth quarter of 2019 (see uncoupling, while exactly the opposite idea turned Figure 6), where service cancellations of approxi- out to be cheaper in the third case. Assessing these mately 360,600 total train kilometers were prevented. concepts would have been impossible without FEO. Extrapolating this to an annual value, FEO allows DB DB Cargo uses FEO for strategic and operational loco- including a timetable improvement, was superior to motive scheduling in their highly demand-driven and operating a diverse locomotive fleet. The new con- Research Campus MODAL: Success Stories volatile business. A particular problem is cross-border cept leads to an increase in efficiency of 3% to 5%, operations, which constitute more than 60% of DB Car- which amounts to savings of 27 million euros per year. go’s transports. These require either special locomo- tives for every country or expensive so-called interop- The Edelman Competition is named after Franz Edel- erable locomotives. The Alps, in particular, can be man, who fled from Germany to England in the late crossed using the Gotthard or the Lötschberg/Simplon 1930s as a teenager, on to Canada, and finally to passes in Switzerland (“over the Alps”), or the Lötsch- the US. He worked for 30 years at the Radio Corpo- berg/Simplon and the new Gotthard Basis tunnels ration of America, where he set up one of the earliest (“under the Alps”). The feasible train compositions industrial OR/MS groups. “However, the most pow- depend on the inclination of the track, the weight and erful incentive will continue to be effective support the length of the train, and the delivery deadline. In of the firm’s managerial and professional activities,” a series of optimization projects, it became clear that said Edelman [E82]. The FEO system achieves exactly an integrated operation of cross-border transports, that for the vehicle-rotation operations at DB.
Figure 7: DB Regio used FEO for successful tendering in Baden-Württemberg (left) and DB Cargo for cross-border freight-traf- fic operations (right). EnergyLab Enters Finals of the INFORMS 2020 Innovative Applications in Analytics Award
The MODAL EnergyLab team from ZIB, TU, LBW The “Innovative Applications in Analytics” award, pre- Optimization, and Open Grid Europe made it to sented by the Analytics Society of INFORMS, recogniz- the finals of the INFORMS 2020 Innovative Ap- es the creative combination of descriptive, predictive, plications in Analytics award. A natural-gas-dis- and prescriptive analytics in innovative applications patching decision-support system, developed to provide insights and/or business value. In 2020, by the team, employs a novel combination of the jury selected six finalists out of 33 contributions analytics to compute smart and forward-looking from across the globe. The MODAL EnergyLab team recommendations for safe and efficient control of with researchers from ZIB and TU, LBW Optimization, one of the largest German natural gas networks. and Open Grid Europe is proud to be one of them. 73 This system is the starting point to an optimized control of the network, which will allow the in- The team presented its work on optimal control of the tegration of new technologies, especially those German gas network with its manifold connections related to hydrogen. to the European network. Open Grid Europe (OGE)
Figure 8: The MODAL team presents its contribution in a video presentation and a video conference with the jury. is a transmission-system operator responsible for the led to a software suite that is in productive use at delivery of more than 25% of the German primary en- OGE for controlling the gas network. ergy consumption and transporting about the same Research Campus MODAL: Success Stories amount to Western and Southern Europe. To achieve For this to work, it was necessary to utilize three types this, they operate a natural-gas network of compa- of analytics: rable length to the German Autobahn. The 24/7 dis- patching center operates 92 compressor units, almost · Descriptive: modeling and simulating the gas 300 regulators, and more than 3,000 valves to control flow in the network. the network to meet all demands. This is a network · Predictive: predicting future gas supply and that has been built in the course of the last 100 years; demand at the entries and exits of the network. the oldest still-active part was built in 1916. There · Prescriptive: recommending network-control is no computer-readable description and it is not measures to ensure safe and efficient operation stuffed full of sensors. This is solid engineering made of the network. of thousands of tons of steel. Thus, the network is operated mainly based on measured data, which is Every 15 minutes, the optimization core computes only available in the “rearview mirror” and the dis- recommendations based on the current state of patcher’s expert knowledge. To improve this situation the network, its past, and its technical capabilities. and to anticipate and prevent critical situations in the A single station in the network can have more than network, the MODAL team has developed a smart, 1,250 discrete operational modes. Each mode may forward-looking, analytics-based decision-support include target values for continuous quantities, such system. This was one of the main research projects as the target pressure of a regulator. First, an hourly of MODAL’s GasLab, the “ancestor” of the EnergyLab forecast for the more than 1,000 entry and exit points in MODAL’s first funding period. It is the main success of the network for the next 24 hours is computed, story of the GasLab that five years of research have employing a mixture of machine learning and optimi- The Combinatorics of Gernsheim
30,000,000,000,000,000 mathematically 1,285 relevant operation possible combinations 200,000 feasible operation modes extracted using of valve and modes identified based on analytical evaluation compressor states. practitioners knowledge. of historical data.
Figure 9: Largest individual compressor station at Gernsheim.
zation on a preselected set of features most suitable the dispatchers on dedicated iPads, providing them 75 for each entry or exit. Based on these forecasts, the with a set of directions. recommended operational measures are computed. However, an exact transient model, including all dis- As the MODAL GasLab has progressed to the En- crete and continuous variables for the full network, ergyLab, integrating new technologies into this sys- is computationally intractable. Therefore, we employ tem has started. On a larger scale, companies have a two-phase approach, decomposing the different started to produce hydrogen from excess power. aspects of the problem by exploiting the gas net- This hydrogen can be injected into the natural-gas work’s topology: a coarse model computes amounts infrastructure. Since the maximum concentration of hy- and directions of flow, computing when to transport drogen anywhere in the network is limited, detailed how much gas on which path through the network. tracking of the gas composition becomes necessary, Then, detailed models for the individual compressor making the network’s operation even more difficult. stations are solved in parallel. The most complex one Also, pure hydrogen transport networks are needed is shown in Figure 9. to fulfill, for example, future industrial and mobility demands. Due to the lower energy density of hydro- The solutions of the detailed models validate and gen, such a network holds only about a quarter of compliment the amounts and directions calculated the line-pack. Thus, velocities need to increase to ful- by the coarse model and compute precise action fill the same demand, and the control becomes much recommendations. The objective is to fulfill demands more dynamic. Employees of the gas dispatchers while minimizing costly mode changes considering cannot scale such increasing complexity and scope many additional constraints needed to ensure the at will. The MODAL EnergyLab navigation system will practical feasibility of the recommended action. Fi- provide solutions for these changing demands of en- nally, the results are summarized and displayed to ergy dispatching. CO@WORK Online 2020
CO@Work is one of the most promi- However, motivated by the COVID-19 cri- nent training and outreach activities sis, the whole 2020 summer school was of MODAL. In 2020, this time in the transferred to a MOOC (massive online form of a massive online open course, open course). it attracted 1,154 participants from 68 countries with a rich program including CO@WORK 2020 had 1,154 participants 50 lectures by 38 lecturers as well as from 68 countries in 17 different time Research Campus MODAL: Success Stories daily hands-on exercise sessions via zones (see Figures 10 and 11). About 54% Jupyter Notebooks running in paral- of participants were PhD students, 22% lel in 1,200 instances in a Kubernetes undergrad students, 15% faculty, and 9% cloud hosted at ZIB. from industry.
The CO@WORK summer school address- To make CO@Work accessible world- es master’s students (in their final year), wide, we made it free of charge and PhD students, postdocs, and everyone undertook significant program changes. else interested in combinatorial optimi- While for on-site workshops, one can zation and mathematical programming have eight straight hours of lecturing in industrial applications. We originally (interleaved with breaks), this is not ex- planned to hold it as a physical event pedient when lecturers and students are with approximately one-hundred partici- spread worldwide. The team at MODAL, pants in Berlin like several times before. particularly the lead organizers Dr. Timo
Figure 10: Distribution of CO@Work participants around the globe. Figure 11: Snapshot of 275 out of more than 1,000 participants.
Berthold (FICO) and Professor Thorsten Koch we made extensive experiments with automatic (ZIB), came up with some unique solutions to extraction and translation of subtitles, as well as transfer the mix of theoretical lectures, panel dis- with using synthesized speech generated from cussions, and hands-on programming experience subtitles. Some courses were then used as a test into an online format. bed to get feedback on the respective user ex- perience. The 50 lectures by 38 lecturers were all prerecord- ed and released as daily content. This included Q&A and the exercise sessions took place as
29 classes from the MODAL partner institutions Zoom webinars. The live programming exercis- 77 Siemens, SAP, Gurobi, FICO, GAMS, FU Berlin, es were accompanied by four virtual breakout and ZIB. This daily video feed enabled the stu- rooms, where tutors helped students interactively dents to watch videos when it suited them and at when they got stuck. We wanted to guarantee their own pace. Twice a day, with an eleven-hour that all students would have ready access to the time shift, we had a 30-minute Q&A session on required software and an identical setup to not the lectures of the previous day and a 90-minute lose any time and motivation through tedious online exercise session. technical support. Therefore, we implemented all exercises via Jupyter Notebooks, which ran inside We uploaded all lectures to a YouTube channel, a Docker image. The IT team at ZIB implemented where they are still available for interested view- a system to run 1,200 instances of this Docker ers (https://www.youtube.com/channel/UCphLz_ image in parallel in a Kubernetes cloud hosted BXrOAInHozAlTsigA/playlists). Additionally, the at ZIB. Each student received their own personal, talks were transferred to bilibili.com, a Chinese identical instance to work in. video-streaming site, as YouTube is blocked in China. The most popular presentations received The summer school was very well received, and more than 2,000 views, some of them still getting we all learned a lot about online teaching. It is more than 100 views per week as of March 2021. quite possible that next time, CO@Work will take To reduce barriers, we produced proper subti- place as a hybrid event to combine the accessi- tles for all 50 lecture videos; around two-thirds bility feats of an online format with the network- of the students reported that they made use of ing possibilities of a physical event. this feature. During the preparation of the course, REFERENCES
Research to Fight COVID-19 [Donati2020] L. Donati, M. Weber, B. G. [Brunton2016a] S.L. Brunton, J.L. Proctor, [1] Sebastian A Müller, Michael Balmer, Keller: Markov Models from the Square and J.N. Kutz: Discovering Governing Equa- Billy Charlton, Ricardo Ewert, Andreas Root Approximation of the Fokker-Planck tions from Data by Sparse Identification of Neumann, Christian Rakow, Tilmann Equation: Calculating the Grid-Dependent Nonlinear Dynamical Systems. Proceed- Schlenther, and Kai Nagel. (2020) Using Flux. Journal of Physics: Condensed Matter, ings of the national academy of sciences, mobile Phone Data for Epidemiological https://doi.org/10.1088/1361-648X/abd5f7, 113(15):3932–3937, 2016. Simulations of Lockdowns: Government 2020 Interventions, Behavioral Changes, and [Brunton2016b] S. L. Brunton, B. W. Brunton, Resulting Changes of Reinfections. medRxiv, [Ernst2020] N. Ernst, K. Fackeldey, A. Vol- J. L. Proctor, and J. Kutz: Koopman Invariant doi:10.1101/2020.07.22.20160093 kamer, O. Opitz, M. Weber: Computation of Subspaces and Finite Linear Representa- Temperature-Dependent Dissociation Rates tions of Nonlinear Dynamical Systems for [2] Hanna Wulkow, Tim Conrad, Natasa of Metastable Protein-Ligand Complexes. Control. PLoS ONE, 11(2):e0150171, 2016. Djurdjevac Conrad, Sebastian A. Mueller, Molecular Simulation, 45(11):904-911, 2019. Kai Nagel, and Christof Schuette (2020) [Klus2018] St. Klus, F. Nüske, P. Koltai, H. Prediction of COVID-19 Spreading and [Rabben2020] R. J. Rabben, S. Ray, M. Wu, I. Kevrekidis, Ch. Schütte, and F. Noé: Optimal Coordination of Counter-Mea- Weber: ISOKANN: Invariant Subspaces Data-Driven Model Reduction and Transfer sures: from Microscopic to Macroscopic of Koopman Operators Learned by a Operator Approximation. Nonlinear Science Models to Pareto Fronts. medRxev, doi: Neural Network. The J. of Chem. Phys., 28(3), pp.985–1010, 2018 10.1101/2020.12.01.20241885 153(11):114109, 2020. [Gelß2019] P. Gelß, St. Klus, J. Eisert, Ch. Understanding and Modeling [Ray2020] S. Ray, V. Sunkara, Ch. Schütte, Schütte: Multidimensional Approximation of
References Complex Biological Systems and M. Weber: How to Calculate pH-De- Nonlinear Dynamical Systems. J. Comput. [Weimann2021] Weimann, K. and Conrad, pendent Binding Rates for Receptor-Ligand Nonlinear Dynam. 14(6), 061006, 2019 T. O. F. (2021) Transfer Learning for ECG Systems Based on Thermodynamic Simula- Classification. Scientific Reports. ISSN 2045- tions with Different Binding Motifs. Molecu- 2322 (In Press) [Wulkow2020] N. Wulkow, P. Koltai, Ch. lar Simulation, Vol. 46, Issue 18, 2020 Schütte: Memory-Based Reduced Model- ing and Data-Based Estimation of Opinion [Melnyk2020] Melnyk, Kateryna and Klus, [Bittracher2020a] Andreas Bittracher and Spreading. Nonlinear Science 31(19), 2020 S. and Montavon, Grègoire and Conrad, Christof Schütte, “A Weak Characterization T. O. F. (2020) Graph Kernel Koopman Em- of Slow Variables in Stochastic Dynamical bedding for Human Microbiome Analysis. [Klus2020] St. Klus, F. Nüske, S. Peitz, J.-H. Systems,” Advances in Dynamics, Optimiza- Applied Network Science, 5 (96) Niemann, C. Clementi, Ch. Schütte: Da- tion and Computation, ed. Oliver Junge et ta-Driven Approximation of the Koopman al., Studies in Systems, Decision and Control Generator: Model Reduction, System Iden- [Rams2020] Rams, Mona and Conrad, T. O. (Springer International Publishing), 132–50, tification, and Control. Physica D: Nonlinear F. (2020) Dictionary Learning for Transcrip- 2020, https://doi.org/10.1007/978-3-030- Phenomena, Volume 406:132416, 2020 tomics Data Reveals Type-Specific Gene 51264-4_6. Modules in a Multi-Class Setting. Informa- tion Technology [Carderera2021] A. Carderera, S. Pokutta, [Bittracher2020b] Andreas Bittracher, Stefan C. Schütte, M. Weiser: CINDy: Conditional Klus, Boumediene Hamzi, Péter Kolta and Gradient-Based Identification of Non-linear [Iravani2019] Iravani, Sahar and Conrad, T. Christof Schütte, “Dimensionality Reduction Dynamics – Noise-Robust Recovery. arX- O. F. (2019) Deep Learning for Proteomics of Complex Metastable Systems via Kernel iv:2101.02630, 2021. Data for Feature Selection and Classifica- Embeddings of Transition Manifolds,” Jour- tion. Lecture Notes in Computer Science, nal of Nonlinear Science, 31, no. 1, 2020, 11713 . https://doi.org/10.1007/s00332-020-09668-z. Wind under Your Wings – Discrete-Continuous Free Flight Planning [Zhang2019] Zhang, W. and Klus, S. and [Bittracher2021] Andreas Bittracher and Oxley, D., & Jain, C. (2015). Chapter 1.4 Conrad, T. O. F. and Schütte, Ch. (2019) Christof Schütte, “A Probabilistic Algo- Global Air Passenger Market: Riding Out- Learning Chemical Reaction Networks from rithm for Aggregating Vastly Undersam- Periods of Turbulence, Travel and Tourism Trajectory Data. SIAM Journal on Applied pled Large Markov Chains,” Physica D: Competitiveness Report 2015. International Dynamical Systems, 18(4). ISSN 1536-0040 Nonlinear Phenomena 416 (February 1, Air Transport Association. World Economic 2021): 132799, https://doi.org/10.1016/j. Forum. physd.2020.132799. C.A. Wells, P.D. Williams, N.K. Nichols, D. (Brunton et al., 2016) Steven. L. Brunton, (Hazan and Luo, 2016) Elad Hazan and Kalise, I. Poll. Reducing transatlanticflight Joshua L. Proctor, and J. Nathan Kutz Haipeng Luo. Variance-Reduced and Pro- emissions by fuel-optimised routing. Env. (2016). Discovering Governing Equations jection-Free Stochastic Optimization. Pro- Res. Letters 16:025002, 2021. from Data by Sparse Identification of Non- ceedings of the 33rd International Confer- linear Dynamical Systems. Proceedings of ence on Machine Learning, pp. 1263–1271, the national academy of sciences, 113(15), 2016. M. Blanco, R. Borndörfer, N.-D. Hoang, A. pp. 3932–3937, 2016. Kaier, A. Schienle, T. Schlechte, and S. Schlo- bach. Solving Time Dependent Shortest (Holloway, 1974) C. A. Holloway. An Ex- Path Problems on Airway Networks Using (Carderera et al. 2021) Alejandro Cardere- tension of the Frank and Wolfe Method of Super-Optimal Wind. 16th Workshop on ra, Sebastian Pokutta, Christof Schütte, and Feasible Directions. Mathematical Program- Algorithmic Approaches for Transportation Martin Weiser (2021). CINDy: Conditional ming, 6, pp. 14–27, Dec. 1974. doi:10.1007/ Modeling, Optimization, and Systems (AT- Gradient-Based Identification of Non-lin- BF01580219. MOS 2016). ear Dynamics – Noise-Robust Recovery. Preprint. (see, e.g., Jaggi, 2013) Martin Jaggi. Re- R. Borndörfer, F. Danecker, and M. Weiser. visiting Frank-Wolfe: Projection-Free Sparse A Discrete-Continuous Algorithm for Free (Combettes et al., 2020) Cyrille W. Convex Optimization. International Confer- Flight Planning. Algorithms 14(1):4, 2021. Combettes, Christoph Spiegel, and ence on Machine Learning, pp. 427-435, Sebastian Pokutta. Projection-Free Adaptive 2013 PMLR. Gradients for Large-Scale Optimization. R. Borndörfer, F. Danecker, and M. Weiser. Preprint, 2020. A Priori Error Bounds for Graph-Based Tra- (Lan and Zhou, 2016) Guanghui Lan and jectory Optimization in Free Flight Planning. Yi Zhou. Conditional Gradient Sliding for In preparation 2021. (Combettes and Pokutta, 2021) Cyrille W. Convex Optimization. SIAM Journal on Op- Combettes and Sebastian Pokutta. Com- timization, 26(2), pp.1379–1409, 2016. plexity of Linear Minimization and Projec-
P. Deuflhard, M. Weiser. Numerische Math- 79 tion on Some Sets. Preprint, 2021. ematik 3: Adaptive Lösung partieller Differ- (Levitin and Polyak, 1966) Evgeny S. Levitin entialgleichungen. de Gruyter 2020. and Boris T. Polyak. Constrained Minimiza- (Defazio and Bottou, 2019) Aaron Defazio tion Methods. USSR Computational mathe- and Leon Bottou. On the Ineffectiveness of matics and mathematical physics 6.5, pp. C.A. Wells, P.D. Williams, N.K. Nichols, D. Variance Reduced Optimization for Deep 1–50, 1966. Kalise, and I. Poll. Reducing Transatlantic Learning. Advances in Neural Information Flight Emissions by Fuel-Optimised Routing. Processing Systems 32, pp. 1755–1765, Env. Res. Letters 16:025002, 2021. (McMahan and Streeter, 2010) H. Brendan 2019. McMahan and Matthew Streeter. Adaptive Bound Optimization for Online Convex Op- IATA, https://www.iata.org/en/iata-repos- (Duchi et al., 2011) John C. Duchi, Elad timization. Proceedings of the 23rd Annual itory/publications/economic-reports/glob- Hazan, and Yoram Singer. Adaptive Sub- Conference on Learning Theory, 2010. al-air-passenger-markets-riding-out-peri- gradient Methods for Online Learning and ods-of-turbulence/ Stochastic Optimization. Journal of Machine (Négiar et al., 2020) Geoffrey Négiar, Gide- Learning Research, 12(61), pp. 2121–2159, on Dresdner, Alicia Tsai, Laurent El Ghaoui, Simple Explanations, Sparsity, 2011. Francesco Locatello, Robert M. Freund, Fa- and Conditional Gradients bian Pedregosa. Stochastic Frank-Wolfe for (Braun et al., 2017) G. Braun, S. Pokutta, (Frank and Wolfe, 1956) Marguerite Frank Constrained Finite-Sum Minimization. Pro- and D. Zink. Lazifying Conditional Gradient and Philip Wolfe. An Algorithm for Quadrat- ceedings of the 37th International Confer- Algorithms. Proceedings of the 34th Inter- ic Programming. Naval Research Logistics ence on Machine Learning, pp. 7253–7262, national Conference on Machine Learning, Quarterly, 3(1-2), pp. 95–110, 1956. 2020. pp. 566–575, 2017.
(Grant et al., 2006) Michael Grant, Stephen (Xie et al., 2020) Jiahao Xie, Zebang Shen, (Braun et al., 2019) G. Braun, S. Pokutta, Boyd, and Yinyu Ye. Disciplined Convex Pro- Chao Zhang, Boyu Wang, and Hui Qian. Ef- D. Tu, and S. Wright. Blended Conditional gramming, Chapter in Global Optimization: ficient Projection-Free Online Methods with Gradients: the Unconditioning of Condition- From Theory to Implementation, L. Liberti Stochastic Recursive Gradient. Proceedings al Gradients. Proceedings of the 36th Inter- and N. Maculan, eds., in the book series of the 34th AAAI Conference on Artificial national Conference on Machine Learning, Nonconvex Optimization and Applications, Intelligence, pp. 6446–6453, 2020. 2019. Springer, pp. 155–210, 2006. (Yurtsever et al, 2019) Alp Yurtsever, Suvrit [BG+17] Ralf Borndörfer, Boris Grimm, [BR+11] Ralf Borndörfer, Markus Reuther, Sra, and Volkan Cevher. Conditional Gradi- Markus Reuther, Thomas Schlechte. Tem- Thomas Schlechte, Steffen Weider. A Hyper- ent Methods via Stochastic Path-Integrated plate-Based Re-optimization of Rolling graph Model for Railway Vehicle Rotation Differential Estimator.Proceedings of the Stock Rotations. Public Transport 9(1-2), P. Planning. 11th Workshop on Algorithmic 36th International Conference on Machine 365–383, 2017 Approaches for Transportation Modeling, Learning, pp. 7282–7291, 2019. Optimization, and Systems, OpenAccess Se- ries in Informatics (OASIcs) (20), P. 146–155, [BK+18] Ralf Borndörfer, Torsten Klug, Leon- 2011. Research Campus MODAL: ardo Lamorgese, Carlo Mannino, Markus Success Stories Reuther, Thomas Schlechte (Eds.). Hand- [B18] Isabel Beckenbach. A Hypergraph book of Optimization in the Railway Indus- [BR+12] Ralf Borndörfer, Markus Reuther, Network Simplex Algorithm. Operations Re- try. Springer International Publishing, 2018. Thomas Schlechte, Steffen Weider.Vehicle search Proceedings 2017, P. 309–316, 2018. Rotation Planning for Intercity Railways. Pro- ceedings of Conference on Advanced Sys- [BM+14] Ralf Borndörfer, Julika Mehrgardt, tems for Public Transport 2012 (CASPT12), [B19] Isabel Beckenbach. Matchings and Markus Reuther, Thomas Schlechte, Kerstin 2012. Flows in Hypergraphs. Dissertation, Freie Waas. Re-Optimization of Rolling Stock Ro- Universität Berlin, 2019. tations. Operations Research Proceedings 2013, P. 49–55, 2014. [BR+15] Ralf Borndörfer, Markus Reuther, Thomas Schlechte, Kerstin Waas, Steffen
References [BB18] Isabel Beckenbach, Ralf Borndörf- Weider. Integrated Optimization of Rolling er. Hall’s and Kőnig’s Theorem in Graphs [BRS14] Ralf Borndörfer, Markus Reuther, Stock Rotations for Intercity Railways. Trans- and Hypergraphs. Discrete Mathematics Thomas Schlechte. A Coarse-to-Fine Ap- portation Science, 50(3), P. 863–877, 2015. 341(10), P. 2753–2761, 2018. proach to the Railway Rolling Stock Rotation Problem. 14th Workshop on Algorithmic Ap- proaches for Transportation Modeling, Opti- [E82] Franz Edelman. Managers, Computer [BGJ10] Ralf Borndörfer, Martin Grötschel, mization, and Systems, OpenAccess Series Systems, and Productivity. Interfaces 12(5), Ulrich Jäger. Planning Problems in Public in Informatics (OASIcs), (42) P. 79–91, 2014. P. 35–46, 1982. Transit. Production Factor Mathematics, acatech und Springer, P. 95–122, 2010. PUBLICATIONS
PEER-REVIEWED (156)
*Samer Alhaddad, Jens *H. Anzt, F. Bach, S. Druskat, Fabian Becker, Natasa Djurdje- Horst-Holger Boltz, Jorge Förstner, Stefan Groth, Daniel F. Löffler, A. Loewe, B. Y. Re- vac Conrad, Raphael A. Eser, Kurchan, Andrea J. Liu (2021). Grünewald, Yevgen Grynko, nard, G. Seemann, A. Struck, Luzie Helfmann, Brigitta Schütt, Fluctuation distributions of Frank Hannig, Tobias Kenter, E. Achhammer, F. Appell, M. Christof Schütte, Johannes Zonk- energy minima in complex Franz-Josef Pfreundt, Chris- Bader, L. Brusch, C. Busse, G. er (2020). The Furnace and the landscapes. Physical Review tian Plessl, Merlind Schotte, Chourdakis, P. W. Dabrowski, Goat—A spatio-temporal model Research, 3(1):013061. https:// Thomas Steinke, Jürgen Teich, P. Ebert, B. Flemisch, S. Friedl, of the fuelwood requirement doi.org/10.1103/PhysRevRe- Martin Weiser, Florian Wende B. Fritzsch, M. D. Funk, V. Gast, for iron metallurgy on Elba search.3.013061 (accepted for (2020). HighPerMeshes – A F. Goth, J.-N. Grad, Sibylle Her- Island, 4th century BCE to 2nd publication 2020-12-18) Domain-Specific Language for mann, F. Hohmann, S. Janosch, century CE. PLOS ONE, 15:1-37. Numerical Algorithms on Un- D. Kutra, J. Linxweiler, T. Muth, https://doi.org/10.1371/journal. Luc Bonnet, Jean-Luc Akian, structured Grids. Euro-Par 2020. Wolfgang Peters-Kottig, F. Rack, pone.0241133 Éric Savin, T. J. Sullivan (2020). (Joint publication: Modeling F. H. C. Raters, S. Rave, G. Re- Adaptive reconstruction of im- and Simulation of Complex Pro- ina, M. Reißig, T. Ropinski, J. Felix Binkowski, Fridtjof Betz, perfectly-observed monotone cesses, Distributed Algorithms Schaarschmidt, H. Seibold, J. Rémi Colom, Martin Ham- functions, with applications to and Supercomputing) (accept- P. Thiele, B. Uekermann, S. merschmidt, Lin Zschiedrich, uncertainty quantification. Al- ed for publication 2020-07-22) Unger, R. Weeber (2020). An Sven Burger (2020). Qua- gorithms, 13(8):196. https://doi. environment for sustainable sinormal mode expansion org/10.3390/a13080196 research software in Germa- Sónia Alves, Rainald Ehrig, Pe- of optical far-field quantities. ny and beyond: current state, ter C. Raffalt, Alwina Bender, Phys. Rev. B, 102:035432. open challenges, and call for *Ralf Borndörfer, Fabian Georg N. Duda, Alison N. Agres https://doi.org/10.1103/Phys- action. F1000Research. https:// Danecker, Martin Weiser (2021).
(2020). Quantifying Asymmetry RevB.102.035432 81 doi.org/10.12688/f1000re- A Discrete-Continuous Algo- in Gait: A Weighted Universal search.23224.1 (Joint publica- rithm for Free Flight Planning. Symmetry Index to Evaluate tion: AI in Society, Science, and Felix Binkowski, Fridtjof Betz, Algorithms, 14(1):4. https://doi. 3D Ground Reaction Forces. Technology, Digital Data and Rémi Colom, Martin Hammer- org/10.3390/a14010004 (Joint Frontiers in Bioengineering and Information for Society, Science, schmidt, Lin Zschiedrich, Sven publication: Modeling and Sim- Biotechnology. (accepted for and Culture) Burger (2020). Modal expan- ulation of Complex Processes, publication 2020-09-18) sion of optical far-field quanti- Network Optimization) (epub ties using quasinormal modes. ahead of print 2020-12-25) Bahareh Banyassady, Luis Bar- N. Anari, N. Haghtalab, S. In EPJ Web Conf., 238:05007. ba, Wolfgang Mulzer (2020). Naor, Sebastian Pokutta, M. https://doi.org/10.1051/epj- Time-Space Trade-Offs for Ralf Borndörfer, Heide Hop- Singh, A. Torrico (2020). Struc- conf/202023805007 Computing Euclidean Minimum pmann, Marika Karbstein, tured Robust Submodular Max- Spanning Trees. Journal of Com- Niels Lindner (2020). Sepa- imization: Offline and Online putational Geometry (JoCG), Felix Binkowski, Lin Zschiedrich, ration of cycle inequalities Algorithms. INFORMS Journal 11(1):525-547. Sven Burger (2020). A Riesz-pro- in periodic timetabling. Dis- on Computing. jection-based method for non- crete Optimization, 100552. linear eigenvalue problems. https://doi.org/10.1016/j. Christiane Becker, Klaus Jäger, J. Comput. Phys., 419:109678. disopt.2019.100552 Sven Burger (2021). Nanopho- https://doi.org/10.1016/j. tonics for Solar Energy. Sol. Ener- jcp.2020.109678 gy Mater. Sol. Cells, 221:110916. Ralf Borndörfer, Niels Lindner, https://doi.org/10.1016/j.sol- Sarah Roth (2020). A Concur- mat.2020.110916 (epub ahead Andreas Bittracher, Christof rent Approach to the Periodic of print 2020-12-15) Schütte (2020). A weak charac- Event Scheduling Problem. terization of slow variables in Journal of Rail Transport Plan- stochastic dynamical systems. ning & Management, 100175. In Oliver Junge, O. Schütze, https://doi.org/10.1016/j.jrt- Gary Froyland, S. Ober-Blo- pm.2019.100175 baum, E. Padberg-Gehle, ed- itors, Advances in Dynamics, Optimization and Computation. Series: Studies in Systems, Deci- sion and Control. 304:132-150. Ralf Borndörfer, Ricardo Euler, Júlia Chaumel, Merlind Schotte, Steffen Christgau, Bettina Zsolt Csizmadia, Timo Berthold Marika Karbstein (2020). Ein Joseph J. Bizzarro, Paul Zaslan- Schnor (2020). Comparing MPI (2020). The confined primal Graphen-basiertes Modell sky, Peter Fratzl, Daniel Baum, Passive Target Synchronization integral: a measure to bench- zur Beschreibung von Preis Mason N. Dean (2020). Co- on a Non-Cache-Coherent mark heuristic MINLP solvers systemen im öffentlichen aligned chondrocytes: Zonal Shared-Memory Processor. Mit- against global MINLP solvers. Nahverkehr. In Volume 002/127 morphological variation and teilungen – Gesellschaft für Infor- Mathematical Programming. of FGSV Heureka, pages 1-15. structured arrangement of matik e. V., Parallel-Algorithmen https://doi.org/10.1007/s10107- cell lacunae in tessellated und Rechnerstrukturen, ISSN 020-01547-5 cartilage. Bone, 134:115264. 0177-0454, 28. PARS-Workshop, Ralf Borndörfer, Thomas Eßer, https://doi.org/10.1016/j. 121-132. Patrick Frankenberger, Andreas Wolfgang Dalitz, Wolfram Sper- bone.2020.115264 (epub Huck, Christoph Jobmann, Boris ber, Hagen Chrapary (2020). ahead of print 2020-02-11) Krostitz, Karsten Kuchenbecker, Steffen Christgau, Thomas Stein- swMATH: A Publication-based Kai Moorhagen, Philipp Nagl, ke (2020). Porting a Legacy Approach to Mathematical Michael Peterson, Markus Re- Fatemeh Chegini, Alena CUDA Stencil Code to oneAPI. Software. SIAM Newsletter, uther, Thilo Schang, Michael Kopanicakova, Rolf Krause, In 2020 IEEE International Par- Volume 53 (Number 06 | July/ Schoch, Hanno Schülldorf, Peter Martin Weiser (2020). Efficient allel and Distributed Processing August 2020). https://doi. Schütz, Tobias Therolf, Kerstin Identification of Scars using Symposium Workshops, IPD- org/10.12752/8009 (accepted Waas, Steffen Weider (2020). Heterogeneous Model Hier- PSW 2020, New Orleans, LA, for publication 2020-05-31) Deutsche Bahn Schedules Train archies. EP Europace. https:// USA, May 18-22, 2020, 359-367. Rotations Using Hypergraph doi.org/10.1093/europace/ https://doi.org/10.1109/IPDP- Jo Devriendt, Ambros Gleixner, Optimization. Informs Journal on euaa402 (accepted for publi- SW50202.2020.00070 Jakob Nordström (2020). Learn Applied Analytics. (accepted for cation 2020-12-09) Publications to Relax: Integrating 0-1 Inte- publication 2020-11-05) Steffen Christgau, Thomas ger Linear Programming with Ying Chen, Nazgul Zakiyeva, Steinke (2020). Leveraging a Pseudo-Boolean Conflict-Driv- Jan Brüning, Thomas Hildeb- Bangzhu Zhu, Thorsten Koch Heterogenous Memory System en Search. In Integration of AI randt, Werner Heppt, Nora (2020). Modeling and Fore- for a Legacy Fortran Code: and OR Techniques in Constraint Schmidt, Hans Lamecker, An- casting the Dynamics of the The Interplay of Storage Class Programming. CPAIOR 2020, gelika Szengel, Natalja Amirid- Natural Gas Transmission Memory, DRAM and OS. In Volume 12296 of LNCS, pages ze, Heiko Ramm, Matthias Network in Germany with the 2020 IEEE/ACM Workshop on xxiv-xxvi. Bindernagel, Stefan Zachow, Demand and Supply Balance Memory Centric High Perfor- Leonid Goubergrits (2020). Constraint. Applied Energy. mance Computing (MCHPC), Jelena Diakonikolas, Alejandro Characterization of the Airflow https://doi.org/10.1016/j.apen- 17-24. https://doi.org/10.1109/ Carderera, Sebastian Pokutta within an Average Geometry of ergy.2020.115597 MCHPC51950.2020.00008 (2020). Locally Accelerated the Healthy Human Nasal Cav- Conditional Gradients. In Pro- ity. Scientific Reports, 3755(10). Ying Chen, Xiuqin Xu, Thor- Rémi Colom, Felix Binkowski, ceedings of AISTATS. https://doi.org/10.1038/s41598- sten Koch (2020). Day-ahead Fridtjof Betz, Martin Ham- 020-60755-3 high-resolution forecasting merschmidt, Lin Zschiedrich, Yannic Ege, Christian Foth, Dan- of natural gas demand and Sven Burger (2020). Qua- iel Baum, Christian S. Wirkner, Jens Buchmann, Bernhard supply in Germany with a hy- si-normal mode expansion Stefan Richter (2020). Making Kaplan, Samuel Powell, Stef- brid model. Applied Energy, as a tool for the design of spherical-harmonics-based fen Prohaska, Jan Laufer 262(114486). https://doi.org/ nanophotonic devices. In Geometric Morphometrics (2020). Quantitative PA to- https://doi.org/10.1016/j.apen- EPJ Web Conf., 238:05008. (SPHARM) approachable for mography of high resolution ergy.2019.114486 https://doi.org/10.1051/epj- 3D images containing large 3-D images: experimental conf/202023805008 cavity openings using Ambient validation in tissue phantoms. Surahit Chewle, Franziska Em- Occlusion – a study using her- Photoacoustics, 17:100157. merling, Marcus Weber (2020). CyrilleW Combettes, Sebastian mit crab claw shape variability. https://doi.org/10.1016/j. Effect of choice of solvent Pokutta (2020). Boosting Frank- Zoomorphology, 139:421-432. pacs.2019.100157 on crystallization pathway of Wolfe by Chasing Gradients. In https://doi.org/10.1007/s00435- Paracetamol: An experimental Proceedings of ICML. 020-00488-z and theoretical case study. Crystals, 10(12):1107. https:// doi.org/10.3390/cryst10121107 Gunar Fabig, Robert Kiewisz, Bernhard Fröhler, Tim Elberfeld, Ambros Gleixner, Daniel Steffy Uwe Gotzes, Kai Hoppmann- Norbert Lindow, James A. Torsten Möller, Hans-Christian (2020). Linear Programming us- Baum (2020). Bounding the Powers, Vanessa Cota, Luis J. Hege, Jan De Beenhouwer, Jan ing Limited-Precision Oracles. final rank during a round Quintanilla, Jan Brugués, Stef- Sijbers, Johann Kastner, Chris- Mathematical Programming, robin tournament with integer fen Prohaska, Diana S. Chu, toph Heinzl (2020). Analysis 183(1-2):525-554. https://doi. programming. Operational Re- Thomas Müller-Reichert (2020). and comparison of algorithms org/10.1007/s10107-019-01444-6 search, 1866-1505. https://doi. Sperm-specific meiotic chro- for the tomographic reconstruc- org/https://doi.org/10.1007/ mosome segregation in C. el- tion of curved fibres. Nonde- s12351-020-00546-w Ambros Gleixner, Gregor Hen- egans. eLife, 9:e50988. https:// structive Testing and Evaluation, del, Gerald Gamrath, Tobias doi.org/10.7554/eLife.50988 35(3):328-341. https://doi.org/1 Achterberg, Michael Bastub- Andreas Griewank, Tom 0.1080/10589759.2020.1774583 be, Timo Berthold, Philipp M. Streubel, Caren Tischendorf Konstantin Fackeldey, Jonas Christophel, Kati Jarck, Thorsten (2020). On the abs-polynomial Röhm, Amir Niknejad, Surahit Gerald Gamrath, Timo Ber- Koch, Jeff Linderoth, Marco Lüb- expansion of piecewise smooth Chewle, Marcus Weber (2021). thold, Domenico Salvagnin becke, Hans Mittelmann, Derya functions. Optimization Methods Analyzing Raman Spectral Data (2020). An exploratory com- Ozyurt, Ted Ralphs, Domenico and Software. https://doi.org/1 without Separabiliy Assumption. putational analysis of dual Salvagnin, Yuji Shinano (2020). 0.1080/10556788.2020.181744 Journal of Mathematical Chem- degeneracy in mixed-integer MIPLIB 2017: Data-Driven Com- 8 (epub ahead of print 2020- istry, 3(59):575-596. https://doi. programming. EURO Journal pilation of the 6th Mixed-Integer 09-01) org/10.1007/s10910-020-01201- on Computational Optimi- Programming Library. Mathe- 7 (accepted for publication zation, 241-246. https://doi. matical Programming Computa- Philipp Gutsche, Xavier Garcia 2020-11-25) org/10.1007/s13675-020-00130-z tion. https://doi.org/10.1007/ Santiago, Philipp-Immanuel s12532-020-00194-3 (accepted Schneider, Kevin McPeak, for publication 2020-09-10) Nando Farchmin, Martin Ham- Patrick Gemander, Wei-Kun Manuel Nieto-Vesperinas, Sven merschmidt, Philipp-Immanuel Chen, Dieter Weninger, Leona Burger (2020). Role of Geo-
Schneider, Matthias Wurm, Gottwald, Ambros Gleixner Christoph Gorgulla, Konstantin metric Shape in Chiral Optics. 83 Bernd Bodermann, Markus Bär, (2020). Two-row and two-column Fackeldey, Gerhard Wagner, Symmetry, 12:158. https://doi. Sebastian Heidenreich (2020). mixed-integer presolve using Haribabu Arthanari (2020). Ac- org/10.3390/sym12010158 Efficient Bayesian inversion for hashing-based pairing meth- counting of Receptor Flexibility shape reconstruction of lithog- ods. EURO Journal on Compu- in Ultra-Large Virtual Screens Sebastian Götschel, An- raphy masks. J. Micro/Nanolith. tational Optimization, 8(3-4):205- with VirtualFlow Using a Grey ton Schiela, Martin Weiser MEMS MOEMS, 19:024001. 240. https://doi.org/10.1007/ Wolf Optimization Method. (2021). Kaskade 7 – a Flex- https://doi.org/10.1117/1. s13675-020-00129-6 Supercomputing Frontiers and ible Finite Element Toolbox. JMM.19.2.024001 Innovations, 7(3):4-12. https:// Computers and Mathematics doi.org/10.14529/jsfi200301 Ambros Gleixner, Stephen with Applications, 81:444-458. Joshua Feis, Dominik Beutel, Maher, Benjamin Müller, https://doi.org/10.1016/j.cam- Julian Köpfler, Xavier Garcia João Pedro Pedroso (2020). Christoph Gorgulla, Andras wa.2020.02.011 (epub ahead Santiago, Carsten Rockstuhl, Price-and-verify: a new al- Boeszoermnyi, Zi-Fu Wang, of print 2020-03-07) Martin Wegener, Ivan Fernan- gorithm for recursive circle Patrick D. Fischer, Paul Coote, dez-Corbaton (2020). Helici- packing using Dantzig–Wolfe Krishna M. Padmanabha Das, Martin Hammerschmidt, San- ty-Preserving Optical Cavity decomposition. Annals of Oper- Yehor S. Malets, Dmytro S. dra Döpking, Sven Burger, Modes for Enhanced Sensing ations Research, 284(2):527-555. Radchenko, Yurii Moroz, David Sebastian Matera (2020). Field of Chiral Molecules. Phys. https://doi.org/10.1007/s10479- A. Scott, Konstantin Fackeldey, Heterogeneities and Their Rev. Lett., 124:033201. https:// 018-3115-5 Moritz Hoffmann, Iryna Iavniuk, Impact on Photocatalysis: doi.org/10.1103/PhysRev- Gerhard Wagner, Haribabu Ar- Combining Optical and Kinet- Lett.124.033201 thanari (2020). An open-source Ambros Gleixner, Nils-Chris- ic Monte Carlo Simulations drug discovery platform en- tian Kempke, Thorsten Koch, on the Nanoscale. J. Phys. ables ultra-large virtual screens. Daniel Rehfeldt, Svenja Uslu Chem. C, 124:3177. https://doi. Nature, 580:663-668. https://doi. (2020). First Experiments with org/10.1021/acs.jpcc.9b11469 org/https://doi.org/10.1038/ Structure-Aware Presolving for s41586-020-2117-z a Parallel Interior-Point Meth- od. In Operations Research Proceedings 2019, 105-111. https://doi.org/10.1007/978-3- 030-48439-2_13 Martin Hanik, Hans-Christian Margaux Hoang, Philippe Klaus Jäger, Peter Tillmann, Eu- Larousse Khosravi Khorashad, Hege, Christoph von Tycowicz Garnier, Jeremie Lasue, Henri gene A. Katz, Christiane Becker Lucas V. Besteiro, Miguel A. (2020). Bi-invariant Two-Sample Rème, Maria Teresa Capria, (2021). Perovskite/silicon tan- Correa-Duarte, Sven Burger, Tests in Lie Groups for Shape Kathrin Altwegg, Matthias dem solar cells: Effect of lumi- Zhiming M. Wang, Alexan- Analysis. In Shape in Medical Läuter, Tobias Kramer, Martin nescent coupling and bifaciali- der O. Govorov (2020). Hot Imaging, 44-54. https://doi. Rubin (2020). Investigating the ty. Sol. RRL, 5:2000628. https:// electrons generated in chiral org/10.1007/978-3-030-61056- Rosetta/RTOF observations of doi.org/10.1002/solr.202000628 plasmonic nanocrystals as a 2_4 comet 67P/Churyumov-Gerasi- (epub ahead of print 2020-12- mechanism for surface photo- menko using a comet nucleus 12) chemistry and chiral growth. model: Influence of dust man- J. Am. Chem. Soc., 142:4193. Martin Hanik, Hans-Christian tle and trapped CO. Astrono- https://doi.org/10.1021/ Hege, Anja Hennemuth, Chris- Klaus Jäger, Peter Tillmann, my & Astrophysics, 638:A106. jacs.9b11124 toph von Tycowicz (2020). Christiane Becker (2020). https://doi.org/10.1051/0004- Nonlinear Regression on Man- Detailed illumination model 6361/201936655 ifolds for Shape Analysis using for bifacial solar cells. Opt. Ilja Klebanov, Ingmar Schuster, Intrinsic Bézier Splines. In Proc. Express, 28:4751. https://doi. T. J. Sullivan (2020). A rigorous Medical Image Computing and Kai Hoppmann-Baum, Gioni org/10.1364/OE.383570 theory of conditional mean Computer Assisted Intervention Mexi, Oleg Burdakov, Carl Jo- embeddings. SIAM Journal (MICCAI), 617-626. https://doi. han Casselgren, Thorsten Koch on Mathematics of Data Sci- Klaus Jäger, Peter Tillmann, org/10.1007/978-3-030-59719- (2020). Minimum Cycle Parti- ence, 2(3):583-606. https://doi. Eugene A. Katz, Christiane 1_60 tion with Length Requirements. org/10.1137/19M1305069 Becker (2020). Simulating bifa- In Emmanuel Hebrard, Nysret cial perovskite/silicon tandem Musliu, editors, Integration of Publications Luzie Helfmann, Natasa Djurd- solar cells in large PV fields. Ilja Klebanov, Alexander Sikor- Constraint Programming, Artifi- jevac Conrad, Ana Djurdjevac, In OSA Advanced Photonics ski, Christof Schütte, Susan- cial Intelligence, and Operations Stefanie Winkelmann, Christof Congress, OSA Technical Di- na Röblitz (2020). Objective Research, Volume 12296 of Schütte (2021). From interacting gest, page PvTh1G.3. https:// priors in the empirical Bayes Lecture Notes in Computer Sci- agents to density-based model- doi.org/10.1364/PVLED.2020. framework. Scandinavian Jor- ence, pages 273-282, Springer ing with stochastic PDEs. Com- PvTh1G.3 nal of Statistics. https://doi. International Publishing. https:// munications in Applied Math- org/10.1111/sjos.12485 doi.org/10.1007/978-3-030- ematics and Computational 58942-4_18 Markus Kantner, Theresa Science, 16(1):1-32. https://doi. Höhne, Thomas Koprucki, Sven Stefan Klus, Feliks Nüske, org/10.2140/camcos.2021.16.1 Burger, Hans-Jürgen Wünsche, Sebastian Peitz, Jan-Hendrik (accepted for publication 2020- Felix Höfling, Siegfried Dietrich Frank Schmidt, Alexander Niemann, Cecilia Clementi, 11-01) (2020). Finite-size corrections Mielke, Uwe Bandelow (2020). Christof Schütte (2020). Da- for the static structure factor of Multi-dimensional modeling ta-driven approximation of the a liquid slab with open bound- Luzie Helfmann, Enric Ribera and simulation of semiconduc- Koopman generator: Model aries. The Journal of Chemical Borrell, Christof Schütte, Peter tor nanophotonic devices. In reduction, system identifica- Physics, 153:054119. https://doi. Koltai (2020). Extending Tran- Michael Kneissl, Andreas Knorr, tion, and control. Physica D, org/10.1063/5.0017923 sition Path Theory: Periodically Stephan Reitzenstein, Axel Hoff- 406. https://doi.org/10.1016/j. Driven and Finite-Time Dynam- mann, editors, Semiconductor physd.2020.132416 (epub ics. Journal of Nonlinear Sci- Carmen Juds, Johannes Nanophotonics, Volume 194 of ahead of print 2020-03-06) ence, 30:3321-3366. https://doi. Schmidt, Michael Weller, Thor- Springer Series in Solid-State Sci- org/https://doi.org/10.1007/ id Lange, Tim Conrad, Hans ences, pages 241-283, Springer. Thorsten Koch, Ying Chen, Kian s00332-020-09652-7 Boerner (2020). Combining https://doi.org/10.1007/978-3- Guan Lim, Xiaofei Xu, Nazgul Phage Display and Next-gen- 030-35656-9_7 Zakiyeva (2020). A review eration Sequencing for Mate- Felix Hennings, Lovis Anderson, study of functional autoregres- rials Sciences: A Case Study Kai Hoppmann-Baum, Mark Hans Kersting, T. J. Sullivan, sive models with application to on Probing Polypropylene Turner, Thorsten Koch (2020). Philipp Hennig (2020). Conver- energy forecasting. WIREs Com- Surfaces. Journal of the Amer- Controlling transient gas flow in gence rates of Gaussian ODE putational Statistics. https://doi. ican Chemical Society. https:// real-world pipeline intersection filters. Statistics and Comput- org/10.1002/wics.1525 doi.org/https://pubs.acs.org/ areas. Optimization and Engi- ing, 30:1791-1816. https://doi. doi/10.1021/jacs.0c03482 neering. https://doi.org/https:// org/10.1007/s11222-020-09972-4 doi.org/10.1007/s11081-020- 09559-y (epub ahead of print 2020-10-03) Tobias Kramer, Mirta Rodrí- Niels Lindner, Christian Lieb- Heinz-Eberhard Mahnke, Tobias Mattes Mollenhauer, Ingmar guez (2020). Effect of disorder chen (2020). Determining all Arlt, Daniel Baum, Hans-Chris- Schuster, Stefan Klus, Christof and polarization sequences on integer vertices of the PESP tian Hege, Felix Herter, Norbert Schütte (2020). Singular Value two-dimensional spectra of light polytope by flipping arcs. In Lindow, Ingo Manke, Tzulia Si- Decomposition of Operators harvesting complexes. Photo- Dennis Huisman, Christos D. opi, Eve Menei, Marc Etienne, on Reproducing Kernel Hil- synthesis Research, 144:147-154. Zaroliagis, editors, 20th Sympo- Verena Lepper (2020). Virtual bert Spaces. In Oliver Junge, https://doi.org/10.1007/s11120- sium on Algorithmic Approach- unfolding of folded papyri. Jour- O. Schütze, Gary Froyland, 019-00699-6 es for Transportation Modeling, nal of Cultural Heritage, 41:264- S. Ober-Blobaum, K. Pad- Optimization, and Systems 269. https://doi.org/10.1016/j. berg-Gehle, editors, Advances (ATMOS 2020), Volume 85 of culher.2019.07.007 om Dynamics, Optimization and Tobias Kramer (2020). Tran- OpenAccess Series in Informat- Computation. Series: Studies in sient capture of electrons in ics (OASIcs), pages 5:1-5:18. Systems, Decision and Control. magnetic fields, or: comets in Phillip Manley, Sebastian Wal- https://doi.org/10.4230/OASIcs. 304:109-131. the restricted three-body prob- de, Sylvia Hagedorn, Martin ATMOS.2020.5 lem. In Journal of Physics: Con- Hammerschmidt, Sven Burg- ference Series, 1612:012019. er, Christiane Becker (2020). Alberto Montefusco, Mark A. https://doi.org/10.1088/1742- Niels Lindner (2020). Hyper- Nanopatterned Sapphire Peletier, Hans Christian Öt- 6596/1612/1/012019 surfaces with defect. Journal Substrates in Deep-UV LEDs: Is tinger (2021). A Framework of Algebra, 555:1-35. https:// there an Optical Benefit?. Opt. of Nonequilibrium Statistical doi.org/10.1016/j.jalge- Express, 28:3619. https://doi. Mechanics. II. Coarse-Grain- Daniel J. Laydon, Vikram Sun- bra.2020.02.022 org/10.1364/OE.379438 ing. Journal of Non-Equilibrium kara, Lies Boelen, Charles R. Thermodynamics, 46(1):15-33. M. Bangham, Becca Asquith https://doi.org/10.1515/jnet- (2020). The relative contribu- Matthias Läuter, Tobias Kramer, Kateryna Melnyk, Grègoire 2020-0069 (accepted for publi- tions of infectious and mitotic Martin Rubin, Kathrin Altwegg Montavon, Stefan Klus, Tim Con- cation 2020-09-14) spread to HTLV-1 persistence. (2020). The gas production of rad (2020). Graph Kernel Koop-
PLOS Computational Biology. 14 species from comet 67P/ man Embedding for Human 85 https://doi.org/10.1371/journal. Churyumov-Gerasimenko Microbiome Analysis. Applied Hassan Mortagy, Swati Gupta, pcbi.1007470 based on DFMS/COPS data Network Science, 5(96). https:// Sebastian Pokutta (2020). Walk- from 2014-2016. Monthly No- doi.org/10.1007/s41109-020- ing in the Shadow: A New Per- tices of the Royal Astronomical 00339-2 spective on Descent Directions Mathias Lemke, Lewin Stein Society, 498(3):3995-4004. for Constrained Minimization. In (2020). Adjoint-Based Identi- https://doi.org/10.1093/mnras/ Proceedings of NeurIPS. fication of Sound Sources for Alexander Mielke, Alberto staa2643 Sound Reinforcement and Montefusco, Mark A. Peletier Source Localization. In Notes on (2021). Exploring families of Paweł Mrowiński, Peter Numerical Fluid Mechanics and Fabian Löbel, Niels Lindner, energy-dissipation landscapes Schnauber, Arsenty Kaganskiy, Multidisciplinary Design, Volume Ralf Borndörfer (2020). The via tilting: three types of EDP Johannes Schall, Sven Burger, 145 of Notes on Numerical Fluid Restricted Modulo Network convergence. Continuum Me- Sven Rodt, Stephan Reitzenstein Mechanics and Multidisciplinary Simplex Method for Integrat- chanics and Thermodynamics. (2020). Directional single-pho- Design, pages 263-278, Spring- ed Periodic Timetabling and https://doi.org/10.1007/s00161- ton emission from determin- er, Cham. https://doi.org/ Passenger Routing. In Janis S. 020-00932-x (accepted for publi- istic quantum dot waveguide https://doi.org/10.1007/978-3- Neufeld, Udo Buscher, Rain- cation 2020-10-01) structures. Phys. Status Solidi 030-52429-6_17 er Lasch, Dominik Möst, Jörn RRL, 14:2000115. https://doi. Schönberger, editors, Oper- org/10.1002/pssr.202000115 Gunther Mohr, Simon J. Al- ations Research Proceedings Han Cheng Lie, T. J. Sullivan, tenburg, Alexander Ulbricht, 2019, 757-763. https://doi.org/ Aretha Teckentrup (2020). Error Philipp Heinrich, Daniel Baum, https://doi.org/10.1007/978-3- bounds for some approximate Christiane Maierhofer, Kai 030-48439-2_92 posterior measures in Bayesian Hilgenberg (2020). In-situ defect inference. Numerical Mathemat- detection in laser powder bed ics and Advanced Applications fusion by using thermography ENUMATH 2019. https://arxiv. and optical tomography – com- org/abs/1911.05669 (accepted parison to computed tomogra- for publication 2020-06-15) phy. Metals, 10(1):103. https:// doi.org/10.3390/met10010103 Anna Musiał, Kinga Żołnacz, Benjamin Müller, Felipe Serra- Anton Pakhomov, Franz Löch- Sebastian Pokutta (2020). Re- Nicole Srocka, Oleh Kravets, no, Ambros Gleixner (2020). ner, Lin Zschiedrich, Sina Sar- starting Algorithms: Sometimes Jan Große, Jacek Olszewski, Using two-dimensional Projec- avi, Martin Hammerschmidt, there is Free Lunch. In Proceed- Krzysztof Poturaj, Grzegorz tions for Stronger Separation Sven Burger, Thomas Pertsch, ings of CPAIOR. Wojcik, Paweł Mergo, Kamil and Propagation of Bilinear Frank Setzpfand (2020). Far- Dybka, Mariusz Dyrkacz, Mi- Terms. SIAM Journal on Op- field polarization signatures of Luc Pronzato, Guillaume Sagnol chal Dlubek, Kristian Lauritsen, timization, 30(2):1339-1365. surface optical nonlinearity in (2021). Removing inessential Andreas Bülter, Philipp-Im- https://doi.org/https://doi. noncentrosymmetric semicon- points in c- and A-optimal de- manuel Schneider, Lin Zschie- org/10.1137/19M1249825 ductors. Sci. Rep., 10:10545. sign. Journal of Statistical Plan- drich, Sven Burger, Sven Rodt, https://doi.org/10.1038/s41598- ning and Inference, 213:233- Wacław Urbanczyk, Grzegorz 020-67186-0 Benjamin Müller, Gonzalo 252. https://doi.org/10.1016/j. Sek, Stephan Reitzenstein Muñoz, Maxime Gasse, Ambros jspi.2020.11.011 (epub ahead (2020). Front Cover: Plug&Play Gleixner, Andrea Lodi, Felipe Marc Pfetsch, Sebastian Pokut- of print 2020-12-08) Fiber-Coupled 73 kHz Sin- Serrano (2020). On Generalized ta (2020). IPBoost – Non-Con- gle-Photon Source Operating Surrogate Duality in Mixed-Inte- vex Boosting via Integer Pro- in the Telecom O-Band (Adv. Robert Julian Rabben, Sourav ger Nonlinear Programming. gramming. In Proceedings of Quantum Technol. 6/2020). Adv. Ray, Marcus Weber (2020). In Integer Programming and ICML. Quantum Technol., 3:2070061. ISOKANN: Invariant subspaces Combinatorial Optimization: https://doi.org/10.1002/ of Koopman operators learned 21th International Conference, qute.202070061 Mika Pflüger, R. Joseph Kline, by a neural network. The IPCO 2020, 322-337. https:// Analia Fernández Herrero, Journal of Chemical Physics, doi.org/10.1007/978-3-030- Martin Hammerschmidt, Victor 153(11):114109. https://doi. Anna Musiał, Kinga Żołnacz, 45771-6_25 Publications Soltwisch, Michael Krumrey org/10.1063/5.0015132 Nicole Srocka, Oleh Kravets, (2020). Extracting dimension- Jan Große, Jacek Olszewski, Esfandiar Nava-Yazdani, al parameters of gratings Krzysztof Poturaj, Grzegorz Elham Ramin, Ksenia Bestu- Hans-Christian Hege, T. J. Sul- produced with self-aligned Wojcik, Paweł Mergo, Dybka zheva, Carina Gargalo, Danial livan, Christoph von Tycowicz multiple patterning using graz- Kamil, Mariusz Dyrkacz, Michal Ramin, Carina Schneider, Pe- (2020). Geodesic Analysis in ing-incidence small-angle x-ray Dlubek, Kristian Lauritsen, An- dram Ramin, Xavier Flores-Alsi- Kendall’s Shape Space with scattering. J. Micro Nanolithogr. dreas Bülter, Philipp-Immanuel na, Maj M. Andersen, Krist V. Epidemiological Applications. MEMS MOEMS, 19:014001. Schneider, Lin Zschiedrich, Sven Gernaey (2021). Incremental Journal of Mathematical Imag- https://doi.org/10.1117/1. Burger, Sven Rodt, Wacław design of water symbiosis net- ing and Vision, 62(4):549-559. JMM.19.1.014001 Urbanczyk, Grzegorz Sek, works with prior knowledge: https://doi.org/10.1007/s10851- Stephan Reitzenstein (2020). The case of an industrial park 020-00945-w Plug&play fibre-coupled 73 kHz Pedro Pimentel, Angelika in Kenya. Science of the Total single-photon source operating Szengel, Moritz Ehlke, Hans Environment, 751. https://doi. in the telecom O-band. Adv. Chris Oates, Jon Cockayne, Lamecker, Stefan Zachow, Lau- org/https://doi.org/10.1016/j. Quantum Technol., 3:2000018. Dennis Prangle, T. J. Sullivan, ra Estacio, Christian Doenitz, scitotenv.2020.141706 (epub https://doi.org/10.1002/ Mark Girolami (2020). Opti- Heiko Ramm (2020). Automated ahead of print 2020-08-26) qute.202000018 mality criteria for probabilistic Virtual Reconstruction of Large numerical methods. In F. J. Skull Defects using Statistical Mona Rams, Tim Conrad Hickernell, P. Kritzer, editors, Shape Models and Generative Jan Möller, Ali Isbilir, Titiwat (2020). Dictionary Learning for Multivariate Algorithms and Adversarial Networks. Jianning Sungkaworn, Brenda Osberg, transcriptomics data reveals Information-Based Complex- Li, Jan Egger, editors, Towards Christos Karathanasis, Vikram type-specific gene modules ity, 27:65-88. https://doi. the Automatization of Cranial Sunkara, Eugene O Grushevsky, in a multi-class setting. it – In- org/10.1515/9783110635461- Implant Design in Cranioplasty, Andreas Bock, Paolo Annibale, formation Technology, 62(3-4). 005 12439. (accepted for publica- Mike Heilemann, Christof https://doi.org/https://doi. tion 2020-09-21) Schütte, Martin J. Lohse (2020). org/10.1515/itit-2019-0048 Single molecule mu-opioid Mohamed Omari, Alexander receptor membrane-dynamics Lange, Julia Plöntzke, Susanna Sebastian Pokutta, M. Singh, reveal agonist-specific dimer Röblitz (2020). Model-based ex- A. Torrico (2020). On the Un- formation with super-resolved ploration of the impact of glu- reasonable Effectiveness of precision. Nature Chemical Bi- cose metabolism on the estrous the Greedy Algorithm: Greedy ology, 16:946-954. https://doi. cycle dynamics in dairy cows. Adapts to Sharpness. In Pro- org/10.1038/s41589-020-0566-1 Biology Direct, 15. https://doi. ceedings of ICML. org/10.1186/s13062-019-0256-7 Sourav Ray, Vikram Sunkara, Manish Sahu, Ronja Strömsdör- Merlind Schotte, Júlia Chau- Yuji Shinano, N. Tateiwa, S. Na- Christof Schütte, Marcus We- fer, Anirban Mukhopadhyay, mel, Mason N. Dean, Daniel kamura, A. Yoshida, M. Yasuda, ber (2020). How to calculate Stefan Zachow (2020). En- Baum (2020). Image analysis S. Kaji, K. Fujisawa (2020). Mas- pH-dependent binding rates do-Sim2Real: Consistency learn- pipeline for segmentation of sive Parallelization for Finding for receptor-ligand systems ing-based domain adaptation a biological porosity network, Shortest Lattice Vectors Based based on thermodynamic sim- for instrument segmentation. In the lacuno-canalicular system on Ubiquity Generator Frame- ulations with different binding Proc. Medical Image Computing in stingray tesserae. Meth- work. In 2020 SC20: Internation- motifs. Molecular Simulation, and Computer Assisted Interven- odsX, 7:100905. https://doi. al Conference for High Perfor- 46(18):1443-1452. https://doi. tion (MICCAI), Part III, Volume org/10.1016/j.mex.2020.100905 mance Computing, Networking, org/10.1080/08927022.2020.1 12263 of Lecture Notes in Com- (epub ahead of print 2020-05- Storage and Analysis (SC), 834- 839660 puter Science. https://doi.org/ 01) 848. https://doi.org/10.1109/ https://doi.org/10.1007/978-3- SC41405.2020.00064 030-59716-0_75 Sven Rodt, Philipp-Immanuel Philip Scott, Xavier Garcia San- Schneider, Lin Zschiedrich, To- tiago, Dominik Beutel, Carsten Thomas Siefke, Carol Hurtado, bias Heindel, Samir Bounouar, Manish Sahu, Angelika Szengel, Rockstuhl, Martin Wegener, Ivan Johannes Dickmann, Walter Markus Kantner, Thomas Ko- Anirban Mukhopadhyay, Stefan Fernandez-Corbaton (2020). On Dickmann, Tim Käseberg, Jan prucki, Uwe Bandelow, Sven Zachow (2020). Surgical phase enhanced sensing of chiral mol- Meyer, Sven Burger, Uwe Zeit- Burger, Stephan Reitzenstein recognition by learning phase ecules in optical cavities. Appl. ner, Bernd Bodermann, Stefanie (2020). Deterministic Quantum transitions. Current Directions Phys. Rev., 7:041413. https://doi. Kroker (2020). Quasi-bound Devices for Optical Quantum in Biomedical Engineering org/10.1063/5.0025006 states in the continuum for Communication. In Michael (CDBME), 6(1). https://doi. deep subwavelength structural Kneissl, Andreas Knorr, Stephan org/https://doi.org/10.1515/ information retrieval for DUV Felipe Serrano, Robert Schwarz, Reitzenstein, Axel Hoffmann, cdbme-2020-0037 nano-optical polarizers. Opt. Ambros Gleixner (2020). On editors, Semiconductor Nano- Express, 28:23122. https://doi. the relation between the ex- photonics, Volume 194 of org/10.1364/OE.396044 Farouk Salem, Florian Schint- tended supporting hyperplane 87 Springer Series in Solid-State Sci- ke, Thorsten Schütt, Alexander algorithm and Kelley’s cutting ences, pages 285-359, Springer. Reinefeld (2020). Scheduling plane algorithm. Journal of Jan Skrzypczak, Florian Schint- https://doi.org/10.1007/978-3- data streams for low latency Global Optimization, 78:161-179. ke, Thorsten Schütt (2020). RM- 030-35656-9_8 and high throughput on a Cray https://doi.org/10.1007/s10898- WPaxos: Fault-Tolerant In-Place XC40 using Libfabric. Concur- 020-00906-y Consensus Sequences. IEEE Susanne Röhl, Marcus Weber, rency and Computation Prac- Transactions on Parallel and Konstantin Fackeldey (2020). tice and Experience, 32(20):1- Distributed Systems, 31(10):2392- Felipe Serrano, Gonzalo Muñoz Computing the minimal rebind- 14. https://doi.org/10.1002/ 2405. https://doi.org/10.1109/ (2020). Maximal Quadratic-Free ing effect for non-reversible cpe.5563 TPDS.2020.2981891 Sets. In Integer Programming processes. Multiscale Modeling and Combinatorial Optimization: and Simulation. (accepted for Xavier Garcia Santiago, Sven 21th International Conference, Jan Skrzypczak, Florian Schint- publication 2020-11-30) Burger, Carsten Rockstuhl, IPCO 2020, 307-321. https:// ke (2020). Towards Log-Less, Philipp-Immanuel Schneider doi.org/10.1007/978-3-030- Fine-Granular State Machine Ansgar Rössig, Milena Petkovic (2021). Bayesian optimization 45771-6_24 Replication. Datenbank Spek- (2020). Advances in Verifica- with improved scalability and trum, 20(3):231-241. https:// tion of ReLU Neural Networks. derivative information for effi- doi.org/10.1007/s13222-020- Liping Shi, Andrey B. Evlyukhin, Journal of Global Optimization. cient design of nanophotonic 00358-4 Carsten Reinhardt, Ihar Babush- https://doi.org/10.1007/s10898- structures. J. Light. Technol., kin, Vladimir A. Zenin, Sven 020-00949-1 39:167. https://doi.org/10.1109/ Burger, Radu Malureanu, Boris Ana Prates Soares, Daniel JLT.2020.3023450 (epub ahead N. Chichkov, Uwe Morgner, Mi- Baum, Bernhard Hesse, An- of print 2020-09-11) lutin Kovacev (2020). Progres- dreas Kupsch, Bernd Müller, sive Self-Boosting Anapole-En- Paul Zaslansky (2020). Scatter- Gerhard Scholtz, David Knötel, hanced Deep-Ultraviolet Third ing and phase-contrast X-ray Daniel Baum (2020). D’Arcy W. Harmonic During Few-Cycle methods reveal damage to Thompson’s Cartesian transfor- Laser Radiation. ACS Photonics, glass fibers in endodontic mations: a critical evaluation. 7:1655. https://doi.org/10.1021/ posts following dental bur Zoomorphology, 139:293-308. acsphotonics.0c00753 trimming. Dental Materials. https://doi.org/10.1007/s00435- https://doi.org/10.1016/j.den- 020-00494-1 tal.2020.10.018 (epub ahead of print 2020-12-11) Boro Sofranac, Ambros Gleix- Tom Streubel, Caren Tischen- Christoph von Tycowicz (2020). Selina Weiss, Carl Martin ner, Sebastian Pokutta (2020). dorf, Andreas Griewank (2020). Towards Shape-based Knee Grewe, Sally Olderbak, Benja- Accelerating Domain Propaga- Piecewise Polynomial Taylor Osteoarthritis Classification min Goecke, Laura Kaltwasser, tion: an Efficient GPU-Parallel Expansions – The Generaliza- using Graph Convolutional Andrea Hildebrandt (2020). Algorithm over Sparse Matri- tion of Faà di Bruno’s Formula. Networks. In 2020 IEEE 17th Symmetric or not? A holistic ces. In Proceedings of IA^3 at Modeling, Simulation and Opti- International Symposium on approach to the measurement SC20. mization of Complex Processes Biomedical Imaging (ISBI of fluctuating asymmetry from HPSC 2018, 63-82. https://doi. 2020). https://doi.org/10.1109/ facial photographs. Person- org/10.1007/978-3-030-55240- ISBI45749.2020.9098687 ality and Individual Differenc- Lewin Stein, Florian Straube, 4_3 es, 166:1-12. https://doi.org/ Stefan Weinzierl, Mathias Lem- https://doi.org/10.1016/j. ke (2020). Directional sound Roya Ebrahimi Viand, Felix Hö- paid.2020.110137 source modeling using the Yahui Sun, Daniel Rehfeldt, fling, Rupert Klein, Luigi Delle adjoint Euler equations in a Marcus Brazil, Doreen Thomas, Site (2020). Theory and simula- finite-difference time-domain Saman Halgamuge (2020). tion of open systems out of equi- Daniel Werdehausen, Sven approach. Acoustical Society A Physarum-Inspired Algo- librium. The Journal of Chemical Burger, Isabelle Staude, Thom- of America, 148(5):3075-3085. rithm for Minimum-Cost Relay Physics, 153:101102. https://doi. as Pertsch, Manuel Decker https://doi.org/https://doi. Node Placement in Wireless org/10.1063/5.0014065 (2020). General design for- org/10.1121/10.0002425 Sensor Networks. IEEE/ACM malism for highly efficient flat Transactions on Networking. optics for broadband appli- José Villatoro, Martin Zühlke, https://doi.org/10.1109/ cations. Opt. Express, 28:6452. *Ralph L. Stoop, Arthur Straube, Daniel Riebe, Toralf Beitz, TNET.2020.2971770 https://doi.org/10.1364/ Tom H. Johansen, Pietro Tierno Marcus Weber, Hans-Gerd OE.386573 Publications (2020). Collective directional Löhmannsröben (2020). locking of colloidal monolayers Alexander Tesch (2020). A Poly- Sub-ambient pressure IR-MAL- on a periodic substrate. Phys. hedral Study of Event-Based DI ion mobility spectrometer Daniel Werdehausen, Xavier Rev. Lett., 124:058002. https:// Models for the Resource-Con- for the determination of low Garcia Santiago, Sven Burg- doi.org/10.1103/PhysRev- strained Project Scheduling and high field mobilities. Ana- er, Isabelle Staude, Thomas Lett.124.058002 (Joint publica- Problem. Journal of Scheduling. lytical and Bioanalytical Chem- Pertsch, Carsten Rockstuhl, tion: Modeling and Simulation istry, 412:5247-5260. https:// Manuel Decker (2020). Mod- of Complex Processes, Distrib- doi.org/10.1007/s00216-020- eling Optical Materials at the Peter Tillmann, Klaus Jäger, uted Algorithms and Supercom- 02735-0 Single Scatterer Level: The Christiane Becker (2020). Min- puting) Transition from Homogeneous imising levelised cost of elec- to Heterogeneous Materials. tricity of bifacial solar panel Marcus Weber, Weitere Autoren Adv. Theory Simul., 3:2000192. *Arthur Straube, Bartosz G. arrays using Bayesian optimi- (2020). DIN SPEC 2343: Über- https://doi.org/10.1002/ Kowalik, Roland R. Netz, Felix sation. Sustain. Energy Fuels, tragung von sprachbasierten adts.202000192 Höfling (2020). Rapid onset 4:254. https://doi.org/10.1039/ Daten zwischen Künstlichen of molecular friction in liquids C9SE00750D Intelligenzen – Festlegung von bridging between the atomis- Parametern und Formaten. Daniel Werdehausen, Sven tic and hydrodynamic pictures. Christine Reichardt, editor, Burger, Isabelle Staude, Thom- Philipp Tockhorn, Johannes Commun. Phys., 3:126. https:// Beuth Verlag. as Pertsch, Manuel Decker Sutter, Rémi Colom, Lukas doi.org/10.1038/s42005-020- (2020). Flat optics in high Kegelmann, Amran Al-Ashouri, 0389-0 (Joint publication: numerical aperture broad- Marcel Roß, Klaus Jäger, Thom- Marie-Christin Weber, Lisa Modeling and Simulation of band imaging systems. J. as Unold, Sven Burger, Steve Fischer, Alexandra Damerau, Complex Processes, Distribut- Opt., 22:065607. https://doi. Albrecht, Christiane Becker Igor Ponomarev, Moritz Pfeiffen- ed Algorithms and Supercom- org/10.1088/2040-8986/ab8ea2 (2020). Improved Quantum berger, Timo Gaber, Sebastian puting) Efficiency by Advanced Light Götschel, Jens Lang, Susanna Management in Nanotextured Röblitz, Frank Buttgereit, Ra- Solution-Processed Perovskite inald Ehrig, Annemarie Lang Solar Cells. ACS Photonics, (2020). Macroscale mesen- 7:2589. https://doi.org/10.1021/ chymal condensation to study acsphotonics.0c00935 cytokine-driven cellular and matrix-related changes during cartilage degradation. Bio- fabrication, 12(4). https://doi. org/10.1088/1758-5090/aba08f Maraike Willsch, Frank Fried- Jakob Witzig, Ambros Gleixner Christian Würth, Phillip Man- Ralf F. Ziesche, Tobias Arlt, rich, Daniel Baum, Ivo Jurisch, (2020). Conflict-Driven Heuristics ley, Robert Voigt, Doguscan Donal P. Finegan, Thomas M.M. Michael Ohl (2020). A com- for Mixed Integer Programming. Ahiboz, Christiane Becker, Ute Heenan, Alessandro Tengattini, parative description of the INFORMS Journal on Comput- Resch-Genger (2020). Meta- Daniel Baum, Nikolay Kardji- mesosomal musculature in ing. https://doi.org/10.1287/ surface enhanced sensitized lov, Henning Markötter, Ingo Sphecidae and Ampulicidae ijoc.2020.0973 (epub ahead of photon upconversion: towards Manke, Winfried Kockelmann, (Hymenoptera, Apoidea) using print 2020-10-08) highly efficient low power -up Dan J.L. Brett, Paul R. Shearing 3D techniques. Deutsche Ento- conversion applications and (2020). 4D imaging of lithi- mologische Zeitschrift, 67(1):51- nano-scale E-field sensors. um-batteries using correlative Niklas Wulkow, Péter Koltai, 67. https://doi.org/10.3897/ Nano Lett., 20:6682. https:// neutron and X-ray tomography Christof Schütte (2020). Mem- dez.67.49493 doi.org/10.1021/acs.nano- with a virtual unrolling tech- ory-based reduced modeling lett.0c02548 nique. Nature Communications, and data-based estimation of 11:777. https://doi.org/10.1038/ Jakob Witzig, Timo Berthold opinion spreading. Journal of s41467-019-13943-3 (2020). Conflict-Free Learning Nonlinear Science. (accepted Nazgul Zakiyeva, X. Xu (2020). for Mixed Integer Program- for publication 2020-12-12) Nonlinear network autoregres- ming. In Integration of AI and sive model with application to Hans Christian Öttinger, Alber- OR Techniques in Constraint natural gas network forecast- to Montefusco, Mark A. Peletier *Hanna Wulkow, Tim Conrad, Programming. CPAIOR 2020, ing. Mathematics Japonica. (2021). A Framework of Non- Natasa Djurdjevac Conrad, LNCS, pages 521-530. https:// (accepted for publication 2020- equilibrium Statistical Mechan- Sebastian A. Müller, Kai Nagel, doi.org/10.1007/978-3-030- 06-29) ics. I. Role and Types of Fluctua- Christof Schütte (2020). Predic- 58942-4_34 tions. Journal of Non-Equilibrium tion of COVID-19 spreading Thermodynamics, 46(1):1-13. and optimal coordination of https://doi.org/10.1515/jnet- Jakob Witzig, Timo Berthold counter-measures: From micro- 2020-0068 (accepted for publi- (2020). Conflict Analysis for scopic to macroscopic models cation 2020-08-13) MINLP. INFORMS Journal on to Pareto fronts. PLOS One. 89 Computing. https://doi.org/https://doi.org/ 10.1101/2020.12.01.20241885 (Joint publication: Modeling and Simulation of Complex Pro- cesses, Visual and Data-centric Computing)
BOOKS (4)
Peter Deuflhard, Martin Weiser Natalia Selini Hadjidimitriou, Bernhard Reuter (2020). Gener- Stefanie Winkelmann, Christof (2020). Numerische Mathematik Antonio Frangioni, Andrea alisierte Markov-Modellierung. Schütte (2020). Stochastic 3. Adaptive Lösung partieller Lodi, Thorsten Koch (2020). Springer Spektrum, Wiesbaden. Dynamics in Computational Differentialgleichungen. de Mathematical Optimization for Biology. Springer International Gruyter. Efficient and Robust Energy Publishing. Networks. Springer.
DISSERTATIONS (5)
Moritz Ehlke (2020). 3D Recon- Nico Kruber (2020). Approxi- Xavier Garcia Santiago (2020). Narendra Lagumaddepalli struction of Anatomical Struc- mate Distributed Set Reconcil- Numerical methods for shape Venkatareddy (2020). Revealing tures from 2D X-ray Images. iation with Defined Accuracy. optimization of photonic nano- secrets of mussel-glue mimetic Technische Universität Berlin. Humboldt-Universität zu Berlin. structures. Karlsruher Institut für peptides – From advanced Technologie. NMR to computational process modeling. Humboldt-Universität Neveen Ali Salem Eshtewy zu Berlin. (2020). Mathematical Modeling of Metabolic-Genetic Networks. Freie Universität Berlin. IMPRINT
Published by Konrad-Zuse-Zentrum für Informationstechnik Berlin Takustrasse 7 | 14195 Berlin | Germany
Phone +49-30-841 85-0 Fax +49-30-841 85-125 E-mail [email protected] Imprint www.zib.de
Layout Christiane Weimann Grafik + Produktion www.schnittmarke.biz
Printing Lochmann grafische Produktion GmbH
ISSN 0934-5892
©2020 Konrad-Zuse-Zentrum für Informationstechnik Berlin