Advancing Atomic Modeling: Integration of Computational Clusters & Neural Network Techniques

Gaining knowledge about multi-electron atomic structures will increase the understanding of atoms’ behavior in terms of its electronic structure and potential for chemical bonding. This data is derived by solving the Schrödinger equation for a system with multiple electrons, considering electron-electron repulsion and the shielding effect. the multiconfiguration Dirac-Hartree-Fock (MCDHF) method is an approach implemented to solve the Schrödinger equation. There are several software packets such as GRASP (General Relativistic Atomic Structure Package) that are specifically designed for MCDHF calculations. For multi-electron structures the number of iterations needed increase to hundreds million terms for the multiple configurations which increase the computational requirements to perform the calculations. Our study aims to develop a deep learning methodology, by implementing a neural network (NN) to reduce the memory and time requirements when performing calculations for large systems without hindering the accuracy of the calculations. The NN’s task is to preselect and manage the most relevant configurations out of large basis sets selecting the configurations based on a weight scale from previous atomic structure calculations.