Advanced statistical methods are rapidly impregnating many scientific
fields, offering new perspectives on long-standing problems. In
materials science, data-driven methods are already bearing fruit in
various disciplines, such as hard condensed matter or inorganic
chemistry, as well as soft matter to a smaller extent.
When coupling machine learning to molecular simulations, many problems
of interest display dauntingly-large interpolation spaces, limiting
their immediate application without undesired artifacts (e.g.,
extrapolation). The incorporation of physical information, such as
conserved quantities, symmetries, and constraints, can play a decisive
role in reducing the interpolation space. Conversely, physics can help
determine whether an ML prediction should be trusted, acting as a more
robust alternative to the predictive variance.
In this talk I will show how incorporating physics in ML models for
molecular simulations can help in both directions: as prior and
predictive constraint. Illustrations will include advanced force
fields that span large subsets of chemical space, high-throughput
molecular dynamics of drug-membrane thermodynamics, and automated
dimensionality reduction and clustering for molecular kinetics.