| Abstract Scope |
Discovering new lithium-ion cathode chemistries is slow and costly, requiring iterative synthesis characterization cycles. Density functional theory (DFT) delivers quantitative voltages but cannot scale to combinatorial chemical spaces. We present BatteryMat, a three-tier machine learning and DFT framework that accelerates cathode development. A single-pass Atomistic Line Graph Neural Network (ALIGNN), trained on 7,610 sequential delithiation voltages, predicts average voltage with 0.17 V mean absolute error. Survivors are refined by ALIGNN-FF universal-force-field delithiation profiles. They are then validated by automated PBE+U or optB88-vdW+U supercell DFT, with a corrected lithium-metal reference removing a ∼1 V systematic offset. Benchmarked across olivine, spinel, and layered chemistries (LiFePO4, LiMnPO4, LiMn2O4, Li(Ni,Mn,Co)O2, LiCoO2), DFT reproduces experimental voltages within 0.3 V and capacities within 5% for four of five materials. The pipeline ranks 7,474 candidates in seconds, surfacing polyanion phosphate and fluoride hosts to guide downstream synthesis. A demo is available at atomgpt.org/battery. |