Discontinuous Methods for Accurate, Massively Parallel Quantum : Lithium Ion Interface Dynamics from First Principles Mitchell T. Ong,1 Vincenzo Lordi,1 Erik W. Draeger,1 John E. Pask1 Chao Yang,2 Lin Lin,2 Mathias Jacquelin2

1Lawrence Livermore National Laboratory, 2Lawrence Berkeley National Laboratory

Introduction Project Goals Methods

Lithium Ion Batteries and SEI What’s Known about SEI Formation? • Quantum molecular dynamics (QMD) coupled with density • Spectroscopy reveals SEI functional theory (DFT) in VASP & Qbox composition for: • PBE/GGA exchange–correlation functional – LiPF + Ethylene Carbonate (EC) 6 • Projector augmented wave (PAW) method used in VASP; – LiPF6 + Ethyl Methyl Carbonate (EMC) norm-conserving used in Qbox – LiPF + EC/EMC (3:7) 6 • NVT (canonical ensemble) used for equilibration 1 • Experiments provide no H • VASP: Nose–Hoover thermostat mechanistic details of formation § Understand the reaction mechanisms and dynamics at the • Qbox: Berendsen velocity scaling thermostat • Simulations require quantum anode-electrolyte interface that lead to SEI formation and effects to treat chemical reactivity growth • NVE (microcanonical ensemble) used to collect statistics • Solvation structures determined § Examine the transport properties and solvation structures of • Verlet algorithm using 0.5 fs time step with quantum molecular • Li-ion batteries are key components of consumer electronics 13C Li ions in the bulk electrolyte and at the interface • Future QMD runs will use Discontinuous Galerkin density dynamics, but using small system and electric/hybrid-electric transportation technologies functional theory (DGDFT) code currently in development sizes (<600 atoms) and short § To achieve these goals, we perform massively parallel • Solid-electrolyte interphase (SEI) is a product of electrolyte time scales (<25 ps) quantum molecular dynamics simulations at unprecedented • Local, systematically improvable decomposition; important for stability but limits ion transport 19F time and length scales M. Nie, et al., J. Phys. Chem. C., 2013, 117(3), 1257-1267 • Enables scaling to system sizes of ~10,000+ atoms • Understanding formation and composition of SEI important to K. Leung and J. L. Budzien, PCCP, 2010, 12, 6583-6586 § Enable design of new anode-electrolyte combinations for Ganesh et al., J. Phys. Chem B., 2011, 115, 3085-3090 L. Lin, et al., J. Comput. Phys., 2012, 231(4): 2140-2154 improve performance and safety of Li-ion batteries Ganesh et al., J. Phys. Chem C., 2012, 116, 24476-24481 safe, reliable, high-capacity, high-charge rate batteries L. Lin, et al., J. Comput. Phys., 2012, 231(13): 4515-4529 L. Lin, et al., Phys. Rev. B, 2012, 85(23): 235144

Computational Scaling & Code Improvements Initial Results Next Steps

Bulk Ethylene Carbonate (EC) Qbox Improvements Bulk EC Macroscopic Properties Smaller Bulk Electrolyte System 1) Larger bulk electrolyte systems inspired by experiments Benchmark System Radial Distribution Function • 32 EC + LiPF6 • Different electrolytes and salts 50 • C3H4O3 (88.06 g/mol) C-C • Effect of salt concentration and temperature C-H • LiPF6 (151.91 g/mol) • C3H4O3 (88.06 g/mol) 40 • Finite size effects C-O • 318 atoms a • 64 30 • 1102 electrons 2) Anode + Electrolyte • 640 atoms Solid: VASP • Box length: 15.362 Å • Qbox modified to perform self consistent cycle G(R) • Important chemical reactions that lead to SEI formation • 2176 electrons 20 Dotted: Qbox faster using Harris-Foulkes estimate for energy • Density: 1.32 g/cc • Examine the effect of different anode materials • Box length: 19.039 Å convergence 10 • Investigate the effect of liquid environment • Density: 1.36 g/cc Macroscopic Properties • Enhanced performance when including empty 0 • Simulations of ~10,000+ atoms with new DGDFT code 012345 states, important for metallic systems Parameter Varied Cv (J/gŸK) Viscosity (cP) R (Å) Reference (PBE, NVE, 450eV, 310K) 2.96 1.00 Vibrational Density of States Ensemble (NVT) 2.81 1.04 VASP & Qbox Scaling Comparisons Strong Scaling on Blue Gene/Q Van der Waals 3.63 1.51 Exchange-Correlation (LDA) 2.86 1.69 • 98 EC + 6 LiPF6 + Temperature (450K) 2.88 0.31 Summary

4L-Graphite (H- Experiment (EC) 1.52 (Cp) 1.86 (40°C) terminated) 3.5 Ions together 3 § VASP faster for smaller systems, Qbox scales better for • 1700 atoms Li–O Ions separated 2.5 larger systems • 6020 electrons 2

g(R) Li–C • 38.5 x 14.8 x 40.8Å 1.5 § Qbox sped up by factor of three for metallic systems Li-O 1

• 0.83 M LiPF6 0.5 § Excellent scaling of Qbox on Vulcan (BG/Q) system; timings

0 show best performance for 1 MPI task/node using all 1 2 3 4 5 6 a (Å) OpenMP threads • Similar structural properties predicted when varying • Excellent scaling up • Structural and vibrational properties in good § Consistency between VASP & Qbox in structural properties • VASP ideal for small systems agreement with experiment parameters; chemistry needs to be verified to 65536 cores for bulk EC + - • Qbox scales better for larger systems • Small variations due to simulation parameters, • If Li and PF6 together, only 75% as many carbonyl • 1700 atoms in 40 sec/ + § Separated LiPF results in more carbonyl oxygen atoms • Qbox utilizes hardware threading & SIMD including and thermostat oxygens around Li compared to dissociated LiPF6 6 MD step around the Li+ ion in the first solvation shell than when cation registers on Vulcan (Blue Gene/Q) unlike VASP 4 MPI Tasks/Node • Li+ diffusivity 2-3x faster when Li+ and PF - separate 16 Threads/Node • Consistency between codes important if 6 and anion are together • Run 10-40 ps/month • Qbox uses algorithms designed for massive advantageous to use one code over another for • Heat capacity not affected by mixing solvents, but § Li+ diffusivity 2-3x faster when LiPF dissociates in solvent scalability (e.g. preconditioned steepest descent) different systems/properties viscosity is nonlinear function of mixing 6

This work performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344, supported by the Department of Energy offices of Basic Energy Sciences (BES) and Advanced Scientific Computing Research (ASCR) through the Scientific Discovery through Advanced Computing (SciDAC) program. LLNL-POST-639130