BatteryMat
ML-accelerated DFT screening of battery cathodes
BatteryMat is a computational pipeline for designing lithium-ion (and beyond-lithium) battery electrodes. It moves candidates through a funnel: high-throughput screening of tens of thousands of structures across multiple working ions (Li, Na, K, Ca, Mg, Al, Zn), voltage and thermodynamic-stability filtering, composite-score ranking of cathode candidates, and then sequential DFT supercell delithiation to compute voltage curves and convex-hull stability. It also estimates ion-migration kinetics via nudged elastic band (NEB) barriers and benchmarks against known materials such as LFP, LMO, LCO, and NMC.
A parallel effort scales the ML screening stage to millions of structures (Alexandria database, 14 working ions). The methods combine VASP DFT, ML interatomic potentials (ALIGNN-FF), and automated thermodynamic analysis.
This work is described in BatteryMat: A Hierarchical Machine-Learning and DFT Framework for Average-Voltage Screening of Lithium-Ion Cathode Materials (arXiv:2607.06645). The three-tier pipeline screens with ALIGNN, validates with a machine-learning force field, and finishes with DFT using automatically selected exchange-correlation functionals. It reproduces commercial lithium-ion voltages within 0.3 V and volumetric capacity within 5 percent, and it flags 71 JARVIS-DFT and 213 Alexandria structures as new cathode candidates.
Presented as a poster at the NIST Quantum Matters in Materials Science (QMMS) Workshop (Gaithersburg, MD).
Demo: atomgpt.org/battery