LLM-Inspired Architecture Models Molecular Potential Energy Surfaces

Researchers have designed an algorithm inspired by large language models (LLMs) to compute high-dimensional potential energy surfaces for molecular systems. The approach represents a molecular system as a graph with nodes, edges, and faces, using interactions between these subsets to construct the energy surface for a system with 51 nuclear dimensions. A family of neural networks associated with graph-theoretic subsystems is employed. The work, published on arXiv (2412.03831), aims to address challenges in computational chemistry, including the prediction of reaction rates.