Physics-based simulation models like molecular dynamics provide valuable insight into complex systems. Using physics-based models in new systems, however, requires significant effort in model parameterization and validation. In this talk, I will show how maximum entropy methods can enable rapid construction of accurate physics-based models by fitting approximately valid models with empirical data, typically gathered from an experiment. This enables physics-based models to be a sort of "lens" for interpreting experimental data by creating physically valid simulations that match data gathered from an experiment. We've developed a maximum entropy approach called Experiment Directed Simulation (EDS) to do this in computational chemistry. EDS has been used for a number of systems in chemistry, including designing antibiotics, designing peptide-based biomaterials, modeling quantum dynamics of water, and modeling lithium electrolyte systems. EDS can work with multiple types and amounts of empirical data, including scalars, vectors, and probability distributions. EDS lacks predictive power without empirical data and is often best-suited as a tool for gaining insight and interpreting empirical data within the context of a physics-based model. This talk will review the EDS method, itsapplications, and discuss how maximum entropy methods provide a bridge between purely empirical models with physics-based models.
Back to Workshop II: Interpretable Learning in Physical Sciences