Jeffrey Law1,Shubham Pandey2,Prashun Gorai2,1,Peter St. John1
National Renewable Energy Lab1,Colorado School of Mines2
Jeffrey Law1,Shubham Pandey2,Prashun Gorai2,1,Peter St. John1
National Renewable Energy Lab1,Colorado School of Mines2
Solid-state batteries (SSBs) are safer, more efficient, and potentially more recyclable than traditional batteries. However, major challenges such as interfacial instabilities remain to be overcome. Well-known solid electrolytes are unstable at the interface with metal anodes (e.g. Li metal) as well as with a high-voltage cathode. While large-scale DFT calculations has enabled functional materials discovery, only a small fraction of the plausible compositions (10<sup>6</sup>-10<sup>9</sup>) can be explored with these relatively expensive computations.<br/> <br/>A central problem in using AI to accelerate the search for novel crystalline materials is finding thermodynamically stable structures. Composition-only models lack the accuracy required for assessing stability, especially for polymorphic structures (<i>1</i>). Graph neural network (GNN) models can achieve impressive accuracy, on the order of 1 kcal/mol, in predicting formation enthalpy and energy above the hull; however, the exact input structure is known only after performing expensive DFT relaxation (<i>2</i>, <i>3</i>). Some progress has been made in predicting stability directly from the unrelaxed structures (<i>4</i>), but the accuracy of these approaches is limited by the degree to which the crystal structure changes during DFT relaxation. In developing models intended to rank the stability of potentially unstable crystal structures, including high-energy hypothetical structures in the training data is critical to achieving reasonable performance (<i>5</i>). In an analogous fashion, training only on fully relaxed energies may bias GNN models to predict lower energies for unrelaxed inputs in high-energy configurations.<br/> <br/>Here we develop a reinforcement learning (RL) framework for designing new crystal structures and apply it to the challenge of discovering stable SSB materials. For the surrogate model, we trained a crystal-graph convolutional neural network (CGCNN) on a carefully designed dataset of close to 150,000 final geometries and total energies of three main structure types: ground-state (ICSD), hypothetical battery structures fully relaxed with DFT, and example hypothetical battery structures where only the cell volume is optimized. In this manner, our surrogate model learns to predict a scale-invariant formation energy of the crystal structure in its given coordination, which we subsequently minimize with RL. Validating our top-performing candidates with full DFT relaxations confirms a high percentage of them to be stable while exhibiting large electrochemical stability windows.<br/> <br/>1. C. J. Bartel <i>et al.</i>, <i>npj Comput. Mater.</i> <b>6</b>, 1–11 (2020).<br/>2. T. Xie, J. C. Grossman, <i>Phys. Rev. Lett.</i> <b>120</b>, 145301 (2018).<br/>3. C. W. Park, C. Wolverton, <i>Phys. Rev. Mater.</i> <b>4</b>, 63801 (2020).<br/>4. K. Pal, C. W. Park, Y. Xia, J. Shen, C. Wolverton, <i>npj Comput. Mater.</i> <b>8</b>, 1–12 (2022).<br/>5. S. Pandey, J. Qu, V. Stevanović, P. St. John, P. Gorai, <i>Patterns</i>. <b>2</b> (2021), doi:10.1016/j.patter.2021.100361.