Application of machine learning to build models of interatomic interaction

Machine learning provides alternative possibilities compared to the classical approach of building physical models. Interatomic interaction models (also known as interatomic potentials or force fields) are an example of an area where machine learning has proved indispensable. In my report, I will tell you how the application of machine learning ideas allows you to build models that are comparable in accuracy to quaternomechanical models, but are several orders of magnitude superior in performance. I will also show how the ideas of active learning (also known as "optimal experiment planning") allow you to automate the process of constructing interatomic potentials, which opens the way to the creation of autonomous algorithms that automate some actions, for example, convergence analysis of computational methods that traditionally should be performed by algorithm users.