Configuration interaction trained by neural networks: Application to model polyaromatic hydrocarbons

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning

Background:

• Configuration interaction (CI) methods are computationally intensive due to large matrix diagonalization.
• Determining the importance of electronic configurations is a key bottleneck in CI calculations.
• Improving computational efficiency is crucial for studying complex molecular systems.

Purpose of the Study:

• To investigate the use of artificial neural networks (ANNs) for learning configuration interaction coefficients.
• To enhance the computational efficiency of stochastic and deterministic CI methods.
• To accurately calculate ground state energies of molecular and model systems.

Main Methods:

• Trained ANNs using configuration interaction coefficients from Full CI and Monte Carlo CI methods.
• Explored various input descriptors and output targets for efficient ANN training.
• Applied trained ANNs to Heisenberg spin chains and ladder systems.

Main Results:

• Achieved excellent training efficiency with the developed ANNs.
• The trained models accurately calculated variational ground state energies.
• Demonstrated the potential of ANNs in approximating CI coefficients.

Conclusions:

• ANNs offer a promising computational shortcut for CI calculations.
• This machine learning approach can significantly reduce the computational cost of quantum chemistry simulations.
• The method is effective for approximating ground state energies in relevant model systems.