GMMLP: An Efficient Software for Searching the Global-Minimum of Clusters Accelerated by Using the Machine Learning...

Searching for the global-minimum (GM) structure of clusters is a fundamental challenge in computational chemistry, as the potential energy surface (PES) of clusters exhibits a vast number of local minima that increase exponentially with cluster size. This work presents GMMLP (Global-Minimum Search of Clusters Accelerated by Machine Learning Potentials), an efficient software package developed for identifying the GM structures of clusters. GMMLP integrates the atom-in-molecules neural network potential (AIMNet2) with an improved genetic algorithm (GA), leveraging the high accuracy of AIMNet2 trained at the ωB97M-D3/def2-TZVPP level of theory and the global search capability of the optimized GA. To validate GMMLP, benchmark tests were performed on nine types of clusters, including (CH₂O)_n, (CH₃NH₂)_n, (CH₃OH)_n, (CH₄)_n, (H₂O)_n, (H₂SO₄)_n, (HNO₃)_n, (NH₃)_n, and [CO(NH₂)₂]_n (n = 1-10). Computational results show that GMMLP efficiently explores the PES, searching a total of 9869 isomers across all benchmarked clusters with a total wall time of 39,055.14 s (~10.8 h). The average computational time per isomer ranges from 0.22 s for (CH₄)_n to 10.01 s for (H₂SO₄)_n, demonstrating remarkable efficiency. Additionally, the evolution of relative energy and optimized structures of low-lying isomers are analyzed to illustrate the reliability of the search process. GMMLP provides a powerful tool for cluster research, enabling fast and accurate GM structure identification for a wide range of clusters, which is crucial for understanding cluster properties and their applications in chemistry, materials science, and related fields.