XAI meets chemistry

ML force fields have recently become highly popular since they are considered a useful compromise between classical force fields (quick, less accurate) and DFT (slow, accurate). More and more there is also a growing need to better understand how ML architectures reflect the underlying physics. For this the field XAI - explainable AI may provide the necessary methodology for furthering this understanding. In this talk, I will first introduce XAI and then show some results analyzing a number of common architectures for MLFFs and how much they actually reflect physical intuition.