Improving the accuracy and efficiency of Machine Learning derived interatomic potentials

Abstract

Understanding molecules and molecular interactions has been an active field of research in many fields like materials discovery, catalysis design, biomedicine, and drug design. Molecular dynamics (MD) simulation is one of the tools which facilitates the study of molecules at the atomistic levels and predict their properties. However, there is always a trade-off between the accuracy and computational cost among different MD simulations. Moreover, MD simulations are also limited to the size and complexity of the system. To overcome these limitations, there have been many developments in the research communities where machine learning (ML) methods are used to develop interatomic potentials for molecules. These ML methods also face obstacles to accurately represent molecular systems. In this thesis, we propose a solution to cross the hurdles faced by ML models. We present, the use of enhanced sampling methods to generate training data for developing interatomic potentials. We illustrate the generalizability of ML derived interatomic potentials and show how enhanced sampling methods improve the accuracy of the potentials which can be used for different thermodynamic ranges.