Deep learning, and in general, auto-differentiation frameworks allow expressing many scientific problems as end-to-end learning tasks. Common themes in scientific machine learning involve learning surrogate functions of expensive simulators, sampling complex distributions directly or time-propagation of known or unknown differential equation systems efficiently. We will describe our recent work in applying deep learning surrogates and auto-differentiation techniques in molecular simulations. In particular, we will explore active learning of machine learning potentials with differentiable uncertainty; the use of deep neural network generative models to learn reversible coarse-grained representations of atomic systems; and the application of differentiable simulations for reaction path finding without prior knowledge of collective variables.
Overfitting and lack of generalizability are constant challenges in AI for science. We will discuss the scalability of active learning to practical applications in molecular simulations and the sensitivity of deep learning answers to molecular problems, like the fitting of pair potentials from observables.