STING, guided active learning for Machine Learned Interatomic Potentials examined on Lithium Thiophosphate solid state electrolytes

All solid state batteries, based on lithium thiophosphate (LiPS) solid state electrolytes, offer a promising route to safer, higher energy density storage. Computational characterisation of these materials, and their interfaces with electrode materials, demands accurate and efficient machine learned interatomic potentials (MLIPs). Foundation MLIPs provide attractive performance out of the box, however reliability across novel materials is poorly characterised. In this work, we systematically benchmark seven foundation MLIPs across six LiPS compositions: two established electrolytes, Li3PS4 and Li7P3S11 along with four novel materials from the Li2S-Li3P tie line, Li11P3S, Li5PS, Li7PS2, Li8P2S. We evaluate performance across geometry optimisation, phonon calculations, and molecular dynamics simulations. Foundation MLIPs show impressive stability and qualitative accuracy, but exhibit inconsistent performance across materials and their predicted properties. Phonon second-order force constants predictions are the most discriminating test with MACE-OMAT being the only model that predicts realistic phonons on novel materials. This demonstrates a lack of quantitative accuracy while using foundation MLIPs, necessitating custom MLIPs with system specific data. To develop reliable custom MLIPs, we introduce Strategic Training via Iterative Network Guidance (STING), a light-weight active learning framework that uses committee based uncertainty estimation to iteratively build system specific training datasets. STING incorporates stable foundation MLIP performance to minimise the required volume of DFT calculations. We also introduce the coefficient of standard deviation in the force components as a metric for structure selection, allowing for scaled comparison across temperature ranges. STING trained MLIPs consistently outperform foundation MLIPs across all benchmarks, showing sub 0.1 meV/atom energy RMSEs, and a reduction in force RMSEs by factors of 2-10 meV/Å. We also observe stable second-order force constants for all materials and a reduction in second-order force constants RMSEs by factors of 2-10. We further demonstrate that the STING coefficient of the standard deviation uncertainty metric produces well balanced training data across all training temperatures. These results establish a practical two stage strategy for development in solid state electrolyte research, and lay the groundwork for future interfacial simulations of LiPS electrolytes against Sulphur cathode material.