Machine learning for molecular simulations
Speaker: Igor Poltavskyi (Department of Physics and Materials Science; Faculty of Science, Technology and Medicine; University of Luxembourg)
Title: Machine learning for molecular simulations
Time: Wednesday, 2021.03.10, 10:00 a.m. (CET)
Place: fully virtual (contact Dr. Jakub Lengiewicz to register)
Format: 30 min. presentation + 30 min. discussion
Abstract: In chemistry and physics, the employment of machine learning (ML) methods has a transformative impact, advancing modeling and improving our understanding of complex molecules and materials. Each ML method comprises a mathematically well-defined procedure, and an increasingly larger number of easy-to-use ML packages for modeling atomistic systems are becoming available. While current approaches mainly focus on developing/improving ML models’ architecture, training sets did not get enough attention. However, training sets are keys to the performance of any ML model, determining its applicability range and predictive power. In this talk, I will address an inherent bias of the reference data caused by their nonuniform nature. On examples of ML force fields trained to reproduce the potential energy surface of molecules, I will demonstrate that the commonly employed measures of the quality of ML models, such as root mean square error, do not provide a full picture. Finally, I will show how combining unsupervised and supervised ML methods can effectively widen the applicability range of ML models to the fullest capabilities of the dataset.