Machine Learning-Assisted Local-to-Global Optimization Strategy for Accelerated Molecular Cluster Structure Prediction

We present a machine learning-assisted local-to-global optimization strategy (LOGOS) for predicting stable molecular cluster structures on potential energy surfaces (PESs). This model facilitates the construction of molecular clusters by identifying localized patterns using deep learning, thereby enhancing the structure-building process. LOGOS predicts new plausible binding sites for a parent structure, and the geometry-optimized daughter structure is subsequently generated in the molecular electrostatic potential (MESP)-topography feature space. This approach is benchmarked by building various ground-state clusters of a prototype (CO₂)_<i>n</i> with <i>n</i> < 30, which are well corroborated by the minimum-energy structures reported in the literature. Therefore, the LOGOS approach holds great promise for accelerating the structure search and prediction of large molecular clusters, guided by the learned topology of smaller clusters. Overall, LOGOS provides a practical solution for efficiently identifying energetically favorable cluster structures in a hierarchical manner. The LOGOS approach, leveraging deep learning algorithms, helps to understand complex patterns within molecular clusters and efficiently predict daughter clusters with minimal computational cost and minimal loss of chemical accuracy.