From data to noise to data for mixing physics across temperatures with generative artificial intelligence

Significance While it is tempting to use high-temperature simulations to infer observations about low temperature, it is not always clear how to do so. Here we demonstrate how using generative artificial intelligence we can mix information from simulations conducted at a set of temperatures and generate molecular configurations at any temperature of interest including temperatures at which simulations were never performed. The configurations we generate carry correct Boltzmann weights, and our model minimizes the generation of spurious unphysical configurations. We demonstrate its use here through combining with replica exchange molecular dynamics in a postprocessing framework for sampling peptide and ribonucleic acid. We believe the framework is extensible to generic simulations and experiments for mixing control parameters other than temperature.