Computational Mathematics and Scientific Computing Seminar

Time and Location:

Speaker:

Link:

While generative models for molecular systems have received significant interest recently, they remain limited by the availability of training data and their ability to represent complex, conformational distributions accurately. To address these issues, we have been exploring a framework that leverages classical Gibbs sampling to extend generative models trained on data from tractable simulations to higher-dimensional contexts. I will illustrate this framework with applications to sampling distributions for molecular complexes and path integrals in quantum statistical mechanics. If time permits, I will also discuss the issue of mode collapse in generative models and our attempts to mitigate it.