CIM Seminar with Peter Wirnsberger

Abstract
Recently there has been a surge in using generative models, most notably normalizing flows, to address challenging problems in statistical mechanics either by developing new learned schemes or by addressing shortcomings of existing techniques. Efficient sampling from high-dimensional Boltzmann distributions given an energy function and accurate free energy estimation are two key problems in the field. In this talk we will focus on both topics and present our recent works on targeted free energy estimation, in which we combine free energy perturbation and Bennett’s acceptance ratio method with normalizing flows to obtain powerful estimators [1, 2]. We illustrate the potential of this approach by estimating solid free energies and briefly discuss the limitations of our current model.

[1]  Wirnsberger, Ballard et al., Targeted free energy estimation via learned mappings, J. Chem. Phys. 153, 144112 (2020)

[2]  Wirnsberger, Papamakarios, Ibarz et al., Normalizing flows for atomic solids, arXiv:2111.0869 (2021)

Registration
Register to the event here. The registration closes February 27.