A Deep Learning-Based Framework for Valence Bond Structure Selection and Weight Prediction

The valence bond (VB) theory offers a chemically intuitive, multiconfigurational framework for analyzing bonding, resonance, and reaction mechanisms. However, its broader application has been limited by high computational costs. In this paper, we present DLVB, a deep learning-based framework that integrates the VB theory with graph transformers through a chemically meaningful representation of VB structures. DLVB accurately predicts VB structural weights without the need for ab initio calculations and provides an efficient selected configuration interaction (SCI) scheme for identifying key configurations that enable the construction of compact VB wave functions. The DLVB-based SCI scheme can identify important VB structures from arbitrary structure sets within a given active space, outperforming traditional ionic-order-based selection methods in both accuracy and scalability. This approach offers a new pathway for extending the applicability of the VB theory to the bonding analysis of systems with larger active spaces and increased molecular complexity.