October 2022
Navya Khare received her MS Dual Degree in Computational Natural Sciences. Her research work was supervised by Dr. Prabhakar Bhimalapuram. Here’s a summary of her research work on Towards efficient exploration of complete relevant phase space: Combined Classification and Regression Scheme (CCRS) for on-the-fly collective variable prediction in condensed matter systems:
Numerical simulations of condensed matter systems, for example biomolecular systems, are notorious for slow dynamics and consequent incomplete sampling lead to incorrect inferences on system thermodynamics and kinetics. In this study, we propose a protocol for efficient exploration of {\it relevant phase space} using machine learning techniques. The suggested protocol explicitly performs machine learning in two concurrent parallel stages: (i) map different metastable states into different classes by a classification model with assumption that large jumps that result in anomalous data (to the current class) are indicative of transitions between metastable states and (ii) construct a non-linear map within individual free energy minima (here, a class) by use of regression model that can learn local information. We apply the protocol on alanine tetrapeptide and show a significant increase in sampling in all relevant phase space of the system; specifically, in comparison to the benchmark study using metadynamics using a well designed collective variable. We report sampling of all highly probable metastable states, and a significant increase in the sampling of relevant low probability metastable states, along with the concomitant increase in the sampling of the transitions between all such metastable states. The protocol proposed in the current study will be useful in construction of complete metastable states connectivity graph, thus enabling computation of thermodynamic and kinetic processes of systems with very large degrees of freedom.