New method lets AI speed up battery simulations without the computing costs
Researchers have created a technique to compress powerful AI models into lightweight versions that run faster while maintaining accuracy—a breakthrough for accelerating battery and energy storage development. The approach, tested on zinc chloride solutions used in next-generation batteries, could help manufacturers simulate electrochemical systems at scale without expensive computing infrastructure.
Originaltitel: Foundation model distillation with PiNNAcLe: Application to the ZnCl<sub>2</sub> aqueous electrolyte solution
<p>Aqueous zinc chloride solutions are central to zinc-based electrochemical energy storage and electrochemical sensors, yet their local structure and dynamics remain challenging to characterize without atomistic simulation. Foundation models for materials chemistry can substantially accelerate such simulations, but they often impose significantly higher computational overhead than bespoke machine-learning interatomic potential. Here we bridge these two regimes by developing a workflow that distils a foundation model into our lightweight PiNet2 architecture, enabled by an extension of PiNNAcLe (Pair-wise Interaction Neural Network with Active Learn-on-the-fly). We demonstrate the approach for ZnCl<sub>2</sub> (aq), and show that the framework can seamlessly switch between two foundation models at different density functional theory levels. Structural and dynamical properties from the distilled models have been validated against the foundation models as well as the density functional theory-based molecular dynamics simulations at matched levels of theory. This work provides a model-agnostic route to harnessing foundation models with low-cost bespoke models, enabling routine and scalable simulations of electrolyte systems.</p>