Exploring the Chemical Space of Noncovalent Molecular Clusters Using Automated Cluster Building Algorithm and Neural...

Global minima of molecular clusters are one of the acute and long-standing problems in the computational chemistry domain. It is important to understand the lowest energy geometry isomer of a given compound for further investigation of physical and chemical properties. The exponential growth of local minima with increasing cluster size makes the problem more relevant. The existing ab initio methods are accurate but very expensive in terms of computational resources. It opens the gap to find more efficient and reliable solutions for rapid exploration. On the other side, neural network potentials are the rising alternatives, with numerically accurate and reliable results for the molecular cluster problem. We integrate AIMNET2, a pretrained model, into our TABU-based PyAR interface. Originally, this model was trained mostly on organic molecules, and we are using it for the PES explorations of molecular clusters, which is a neighboring domain but does not explicitly use any training input. We demonstrate the effectiveness of our methodology by exploring standard molecular clusters of water, ammonia, hydrogen peroxide, methanol, and acetic acid, with aggregation numbers ranging from 1-10. This work represents a significant step toward fully automated computational molecular cluster generation, paving the way for accelerated exploration and discovery in this field.