Machine learning approach to configuration interaction

High precision calculations of properties of complex atomic systems are a major challenge due to the exponential scaling of atomic configurations. We have developed a machine learning approach to identify the most important configurations out of a large set of configurations. Our method replaces the configuration interaction calculation for the complete set of configurations with a series of smaller ones performed on an iteratively expanding subset managed by a neural network. This results in significantly reduced computational memory requirements and reduced runtime. We also use this approach to evaluate the importance of the very large set of configurations, inaccessible with direct computation to solve previously intractable problems.