Surrogate models of quantum mechanical calculations have transformed the simulation
of matter at the atomic scale, by dramatically reducing its computational cost.
Machine-learning models, however, offer more than acceleration. By using models
that reflect the physical priors of the problem, and critically analyzing their
performance as a function of the hyperparameters, one can learn much about the
key structural features, and molecular interactions that determine the properties
of a material. I will present some examples of the application of this
"model introspection", and discuss how, more broadly, machine-learning models
can be interpreted in terms of fundamental physical concepts.
Copyright. All Rights Reserved.