Aqvolt: High-Fidelity Machine-Learned Force Fields for Halogenated Solid-State Electrolyte Materials

The escalating global demand for safer, high-energy density batteries underscores the need for advanced solid-state Li-ion technologies. Halide solid-state electrolyte (SSE) materials have recently emerged as a class of highly promising candidates that offer superior ionic mobility, wider electrochemical stability windows, and deformability for robust interfacial contact compared to their oxide and sulfide counterparts [1]. Despite this promise, closely studying these materials at larger scales has been a challenge. Prior screening approaches for SSEs have relied on holistic methods, obtaining conductivity values from the literature [2], creating machine learning models based on experimental data [3], or running Ab Initio Molecular Dynamics simulations on a few promising candidates [4], which are costly and too short to model diffusion on macroscopic time scales. In this work, we present AQVolt: a high-fidelity, off-equilibrium dataset for halogenated SSEs and machine-learned force fields (MLFFs) at the r2SCAN level of theory. Our approach begins with leveraging existing MLFFs [5] to perform machine learning Molecular Dynamics (MLMD) simulations and generating millions of diverse atomic configurations at various temperatures. A stratified sampling and dimensionality reduction methodology [6] is employed to reduce sampling redundancies by 98%. By performing single-point calculations from first-principles on these intelligently sampled structures, we significantly reduce our calculation workload while exploring diverse configurations. Benchmarking on current state-of-the-art foundational potentials suggests that limited exposure to off-equilibrium data can hinder their performance [5, 7-8]. AQVolt-trained MLFFs provide the unprecedented accuracy and speed necessary to conduct more accurate MD simulations, thereby predicting downstream properties pertinent to solid-state electrolytes such as ionic conductivity. AQVolt will accelerate the discovery of next-generation halide SSEs, paving the way for safer and higher-performance solid-state battery materials. References Liu, Z., Chien, P. H., Wang, S., Song, S., Lu, M., Chen, S., Xia, S., Liu, J., Mo, Y., & Chen, H. (2024). Tuning collective anion motion enables superionic conductivity in solid-state halide electrolytes. Nature chemistry, 16(10), 1584–1591. Shen, L., Wang, Z., Xu, S., Law, H. M., Zhou, Y., & Ciucci, F. (2025). Harnessing database-supported high-throughput screening for the design of stable interlayers in halide-based all-solid-state batteries. Nature Communications, 16(1), 3687. Sendek, A. D., Yang, Q., Cubuk, E. D., Duerloo, K. A. N., Cui, Y., & Reed, E. J. (2017). Holistic computational structure screening of more than 12000 candidates for solid lithium-ion conductor materials. Energy & Environmental Science, 10(1), 306-320. Chen, W., Zhao, Y., Zhou, J., Li, S., Lu, C., Zhou, S., ... & Luo, J. (2025). High-throughput screening of halide solid-state electrolytes for all-solid-state Li-ion batteries through structural descriptor. Journal of Alloys and Compounds, 1010, 177167. Kaplan, A. D., Liu, R., Qi, J., Ko, T. W., Deng, B., Riebesell, J., ... & Ong, S. P. (2025). A foundational potential energy surface dataset for materials. arXiv preprint arXiv:2503.04070. Qi, J., Ko, T. W., Wood, B. C., Pham, T. A., & Ong, S. P. (2024). Robust training of machine learning interatomic potentials with dimensionality reduction and stratified sampling. npj Computational Materials, 10(1), 43. Kuner, M. C., Kaplan, A. D., Persson, K. A., Asta, M., & Chrzan, D. C. (2025). MP-ALOE: an r2SCAN dataset for universal machine learning interatomic potentials. npj Computational Materials, 11(1), 352. Huang, X., Deng, B., Zhong, P., Kaplan, A. D., Persson, K. A., & Ceder, G. (2025). Cross-functional transferability in universal machine learning interatomic potentials. arXiv preprint arXiv:2504.05565.