Machine Learning Configuration Interaction for ab Initio Potential Energy Curves.

Area of Science:

• Computational chemistry
• Quantum chemistry
• Machine learning applications

Background:

• Accurate ab initio potential energy curves are crucial for understanding molecular behavior.
• Scalability limitations in traditional methods hinder the computation of large molecular systems.
• Artificial neural networks (ANNs) offer a promising avenue for accelerating quantum chemistry calculations.

Purpose of the Study:

• To develop an enhanced machine learning configuration interaction (MLCI) approach for efficient and accurate ab initio potential energy curve calculations.
• To improve the scalability of MLCI by employing ANNs for duplicate configuration identification.
• To guarantee pure spin states and exploit data transferability in MLCI calculations.

Main Methods:

• Developed an MLCI approach using ANNs for on-the-fly selection of important configurations.
• Integrated ANN as a hash function for efficient duplicate configuration deletion, removing storage barriers.
• Introduced configuration state functions to ensure pure spin states and enable data transferability.
• Applied the enhanced MLCI to calculate potential energy curves for N2, H2O, and CO.

Main Results:

• The enhanced MLCI approach successfully computed accurate ab initio potential energy curves for N2, H2O, and CO.
• MLCI demonstrated lower errors and reduced computational cost compared to stochastic configuration selection for N2 and CO.
• The method proved efficient and scalable, overcoming previous limitations.

Conclusions:

• The improved MLCI method enables efficient and accurate computation of ab initio potential energy curves.
• MLCI offers a significant advancement over existing methods, particularly for systems like N2 and CO.
• This work paves the way for more accessible and accurate quantum chemical calculations using machine learning.