Deep-Learning Approach for the Atomic Configuration Interaction Problem on Large Basis Sets.

Area of Science:

• Computational Physics
• Quantum Chemistry
• Artificial Intelligence

Background:

• High-precision atomic structure calculations necessitate accurate modeling of electronic correlations.
• The configuration interaction (CI) problem, crucial for this modeling, involves computationally expensive multiconfiguration wave function expansions.
• Existing methods face scalability challenges, becoming infeasible even for supercomputers with increasing complexity.

Purpose of the Study:

• To develop a deep-learning approach for preselecting relevant configurations in large CI basis sets.
• To replace computationally intensive full CI calculations with a series of smaller, manageable computations.
• To achieve targeted energy precision in atomic structure calculations efficiently.

Main Methods:

• A deep-learning approach utilizing a convolutional neural network (CNN) was developed.
• The CNN iteratively manages an expanding subset of the CI basis set, selecting the most relevant configurations.
• This approach replaces dense neural network architectures, which were found to be unsuitable for quantum chemistry problems.

Main Results:

• The CNN-based method successfully performs robust and accurate CI calculations.
• The approach was benchmarked on moderate-sized basis sets, validating its accuracy against direct CI calculations.
• The method demonstrated feasibility for prohibitively large CI basis sets where direct computation is impossible.

Conclusions:

• Deep learning, specifically CNNs, offers a viable and efficient solution for the large-scale configuration interaction problem in atomic physics.
• This novel method significantly enhances the feasibility of high-precision atomic structure calculations.
• The approach paves the way for tackling more complex electronic correlation problems in computational chemistry and physics.