Â鶹ÊÓƵ

Synopsis

Machine-learning force fields now reach the accuracy of quantum-mechanical calculations at a cost approaching that of classical force fields, and are becoming a standard tool for molecular simulation. I will give an overview of how these models are built, what they make possible, and current trends. I will then discuss a more theoretical question that tends to get overlooked in favour of chasing fit accuracy: which parts of the interaction are actually determined by the data a model is trained on? I will first discuss this as a general inverse problem and then specialize to non-bonded interactions, which carry much of the physics relevant to biomolecules. There, the answer is subtle because non-bonded interactions are weak compared with the bonded ones that dominate the training data.