Abstract
The generation of equilibrium samples of molecular systems has been along-standing problem in statistical physics. Boltzmann Generators are agenerative machine learning method that addresses this issue by learning atransformation via a normalizing flow from a simple prior distribution to thetarget Boltzmann distribution of interest. Recently, flow matching has beenemployed to train Boltzmann Generators for small molecular systems in Cartesiancoordinates. We extend this work and propose a first framework for BoltzmannGenerators that are transferable across chemical space, such that they predictzero-shot Boltzmann distributions for test molecules without being retrainedfor these systems. These transferable Boltzmann Generators allow approximatesampling from the target distribution of unseen systems, as well as efficientreweighting to the target Boltzmann distribution. The transferability of theproposed framework is evaluated on dipeptides, where we show that itgeneralizes efficiently to unseen systems. Furthermore, we demonstrate that ourproposed architecture enhances the efficiency of Boltzmann Generators trainedon single molecular systems.