The content presents MoreRed, a novel approach for molecular relaxation that reframes the problem as a denoising task using a diffusion model. Unlike classical force field methods and machine learning force field (MLFF) models, MoreRed does not aim to learn the complex physical potential energy surface (PES). Instead, it learns a simpler pseudo PES by treating non-equilibrium molecular structures as noisy versions of their corresponding equilibrium states.
The key technical innovation is the introduction of a diffusion time step predictor, which estimates the appropriate starting point for the reverse diffusion process. This allows MoreRed to handle non-equilibrium structures with arbitrary noise levels, unlike standard diffusion models that require the noise level as an input. MoreRed is shown to outperform classical force fields, semiempirical methods, and MLFF models in terms of structural deviation from reference equilibrium structures and DFT energy levels, while being more robust to distorted inputs.
The authors compare three variants of MoreRed that differ in how they handle the time step prediction. MoreRed-JT, which jointly predicts the time step and the noise using a shared neural network backbone, is found to perform the best. MoreRed requires significantly less training data (only equilibrium structures) compared to MLFF models, which need both equilibrium and non-equilibrium structures with computed energy and force labels.
To Another Language
from source content
arxiv.org
Deeper Inquiries