Sampling thermodynamic ensembles of molecular systems with generative neural networks: Will integrating physics-based models close the generalization gap?

Abstract

If the promise of generative modeling techniques is realized, it may fundamentally change how we carry out molecular simulation. The suite of techniques and models collectively termed “generative AI” includes many different classes of models built for varied types of data, from natural language to images. Recent advances in the machine learning literature that construct ever better generative models, though, do not contend with the challenges unique to complex, molecular systems. To generate a statistically likely molecular configuration, many correlated degrees of freedom must be sampled together, while also satisfying the strong constraints of chemical physics. Recent efforts to develop generative models for biomolecular systems have shown spectacular results in some cases—nevertheless, some simple systems remain out of reach with our present methodology. Arguably, the central concern is data efficiency: we should aim to train models that can meaningfully generalize beyond their training data and hence facilitate discovery. In this review, we discuss methods and future directions for directly incorporating physics-based models into generative neural networks, which we believe is a crucial step for addressing the limitations of the current toolkit.