24 November 2026 - 24 November 2026
1:00PM - 2:00PM
CG218
free
Seminar: "Applications of MACE foundational models on battery electrolytes"
Panagiotis Kourtis
Newcastle University, School of Natural and Environmental Science
Applications of MACE foundational models on battery electrolytes
Foundational Machine Learned Interatomic Potentials (FMLIPs) have recently emerged as a powerful tool for atomistic simulations. MACE, an equivariant graph neural network potential, is one of the most widely used FMLIP architectures [1]. In this talk I will present applications of MACE FMLIPs to study battery electrolytes. The main questions we aim to answer are:
Weak inter-molecular interactions underpin thermodynamic properties such as the density and diffusivity of a system. We demonstrate the out-of-the-box stability of MACE-MP (one of the early MACE FMLIPs) on a Li PF6, EC/EMC mixture - a common battery electrolyte. Fine-tuning the model with as little as 10 configurations strongly improves the model’s inter and intra molecular property predictions. Beyond short-range models, the new charge-aware, MACE-POLAR produces physically relevant partial charges and treats electrostatic interactions across the whole system explicitly via long-range Ewald summation [2]. In an aqueous iron chloride system, this model correctly captures the local reorganization due to different oxidation states of iron, while charge-agnostic models fail. Moreover, we demonstrate that long -range models are important to capture ion-ion correlation effects in concentrated electrolyte systems. We study concentrated aqueous ammonium chloride and ammonium sulphate electrolytes which remain liquid up to high salt concentrations of 6 molkg-1. The OpenFF classical forcefield shows that conductivity including ion-ion correlations estimated using the Einstein-Helfand (EH) formalism converges significantly slower than the typical Nernst-Einstein conductivity. We show that MACE potentials are sufficiently fast to capture nanosecond timescale dynamics at which EH coefficients converge. Long-range models correctly predict the experimentally measured ionic conductivity for both systems, while short-range models struggle with ammonium chloride. We believe the reason for this discrepancy lies within the ability of long-range models to accurately predict polarisation.
[1] Batatia, Ilyes, et al. A foundation model for atomistic materials chemistry. J. Chem. Phys. 163 184110 (2025).
[2] Batatia, Ilyes, et al. "MACE-POLAR-1: A polarisable electrostatic foundation model for molecular chemistry." arXiv preprint arXiv:2602.19411 (2026).
[3] Boothroyd, Simon, et al. "Development and benchmarking of open force field 2.0. 0: the sage small molecule force field." J. Chem. Theor. Comp. 19 3251 (2023).