Unified GNN Architecture Design for High-Throughput Material Screening

Petar Griggs, Lin Li, Rajmonda Caceres MIT Lincoln Laboratory, Lexington, MA Overview Integrated GNN (I-GNN) Architecture Optimization Results Motivation: Approach: Mean Absolute Error • Advanced materials (e.g, materials to withstand extreme environments, • Use graph neural networks (GNN) to extract fundamental features and materials for CO2 capture) play an important role in national security. predict a variety of properties from any new crystal structure Band Gap (eV) Bulk Modulus Shear Modulus • Existing experimental trial-and-error approach is time-consuming and costly. • Build a unified and flexible custom architecture design process to (GPa) (GPa) [2] • AI opportunities, i.e., more data, maturing algorithmic solutions • optimize property-specific GNN architecture CGCNN 0.280 17.567 13.619 Problem: • enable better transfer learning when there is limited training data MEGNET [3] 0.307 16.122 13.426 • Robust feature representation learning of materials to GATGNN [4] 0.395 15.652 12.996 - better capture relationships between a material’s structure and properties I-GNN 0.257 14.665 12.596 - enable high-throughput material screening • Generalizability of extracted crystalline representations to properties different from original training properties when there is limited training data Pearson’s Rho • Methodical exploration of the best GNN architecture for a prediction task Band Gap (eV) Bulk Modulus Shear Modulus (GPa) (GPa) Graph Representations of Crystalline Structures CGCNN [2] 0.943 0.913 0.768 Node features: group number, electronegativity, atomic volume MEGNET [3] 0.917 0.921 0.762 Edge features: bond length, [4] 0.909 0.920 0.772 Global features: temperature, elemental proportions GATGNN Graph Stats: Avg. 28 nodes/unit cell; # of edges increases cubically with radius I-GNN 0.943 0.933 0.778

Abstract Graph Representation Design Space Summary and Future Work Input configurations global feature, radius (1) Form sphere of Conclusion GNN configurations num_gnn, gnn pooling, edge/node/global update specified radius on • Built a unified GNN design space functions, activation, skip connection, aggregation each unit cell node • Develop a unified and flexible I-GNN architecture optimization and create edges with Learning configurations learning rate, weight decay, optimizer all inscribed nodes • Apply it on material property prediction • I-GNN outperforms the existing state-of-the-art GNN architectures (up to 20% improvement) (2) Compress Best Architectures Smaller radii better expanded Future work approximate graph into • Physics-informed representation learning to (1) explore the role of self- covalent bonds unit cell supervised representation learning in supporting better feature priors; • Smaller radii tend to (2) incorporate domain constraints and physics-informed features Larger radii better • Intelligent design of complex material compositions encode crystal’s give better repeating pattern performance References [1] Jiaxuan You, Rex Ying, Jure Leskovec. Design Space for Graph Neural Networks. Neural • Update function is Information Processing Systems (NeurIPS), 2020. [2] Xie, Tian, et al.. "Crystal graph convolutional neural networks for an accurate and property-specific interpretable prediction of material properties." Physical review letters 120.14 (2018): 145301. [3] Chen, Chi, et al. "Graph networks as a universal machine learning framework for molecules 3D Crystal Structure of Compressed Unit and crystals." Chemistry of Materials 31.9 (2019): 3564-3572. Indium Nitride (InN) Cell Representation [4] S.Y. Louis, , et al."Graph convolutional neural networks with global attention for improved materials property prediction." Physical Chemistry Chemical Physics 22, no. 32 (2020)

DISTRIBUTION STATEMENT A. Approved for public release. Distribution is unlimited. This material is based upon work supported by the Under Secretary of Defense for Research and Engineering under Air Force Contract No. FA8702-15-D-0001. Any opinions, findings, conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Under Secretary of Defense for Research and Engineering. © 2021 Massachusetts Institute of Technology. Delivered to the U.S. Government with Unlimited Rights, as defined in DFARS Part 252.227-7013 or 7014 (Feb 2014). Notwithstanding any copyright notice, U.S. Government rights in this work are defined by DFARS 252.227-7013 or DFARS 252.227-7014 as detailed above. Use of this work other than as specifically authorized by the U.S. Government may violate any copyrights that exist in this work.