Molecular relaxation, finding the equilibrium state of a non-equilibrium structure, is an essential component of computational chemistry to understand reactivity. Classical force field methods often rely on insufficient local energy minimization, while neural network force field models require large labeled datasets encompassing both equilibrium and non-equilibrium structures. As a remedy, we propose MoreRed, molecular relaxation by reverse diffusion, a novel purely statistical approach where non-equilibrium structures are seen as noisy instances of the corresponding equilibrium structures. To enable the denoising of arbitrarily noisy inputs using a diffusion model, we further introduce a novel diffusion time step predictor. Notably, MoreRed learns a simpler pseudo potential energy surface instead of the complex physical potential energy surface. It is trained on a significantly smaller and therefore computationally cheaper dataset consisting of solely unlabeled equilibrium structures, avoiding the computation of non-equilibrium structures altogether. We compare MoreRed favourably to classical force fields, equivariant neural network force fields, trained on a large dataset of equilibrium and non-equilibrium data, as well as a semi-empirical tight-binding model. We compare the root-mean-square deviation between the found equilibrium structures and the reference equilibrium structures as well as the energy minimum computed for each of them.
Khaled Kahouli