Regensburg 2025 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 9: Poster
MM 9.21: Poster
Monday, March 17, 2025, 18:30–20:30, P1
GAP vs. MACE: Efficiency evaluation in a liquid electrolyte system — •Anton Beiersdorfer1, Lisa Hetzel1, Carsten Staacke2, Florian Deissenbeck2, and Christopher Stein1 — 1Technische Universität München, München, Germany — 2Cellforce Group GmbH, Tübingen, Germany
Machine learning interatomic potentials (MLIP) have transformed molecular simulations, enabling complex materials to be modeled with increasing accuracy and efficiency. As MLIP models evolve, so does the demand for advanced computing architectures, particularly graphics processing units (GPUs), which can accelerate computations compared to traditional central processing unit (CPU) based systems. However, the high cost associated with GPU resources constrain access in both academia and industry, highlighting the relevance of comparing GPU-based and CPU-based MLIPs under real-world conditions.
To this end two popular MLIPs are examined: the GPU-accelerated MACE model and the CPU-based GAP model applied to a test system of a standard battery electrolyte. The system is selected for its demanding electrostatic interactions in solution, which the MLIPs approximate by learning the local interaction patterns that contribute to the overall electrostatic behavior. Therefore, it represents a significant computational challenge and provides a rigorous benchmark for MLIP accuracy and efficiency. By focusing on these models, the study aims to reveal key differences in computational and numerical performance metrics and resource efficiency as well as in physical performance, particularly through comparisons to experimentally measured properties.
Keywords: Machine learning; Batteries; Liquid electrolyte