Electronic Excited States from Physically Constrained Machine Learning.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Data-driven methods are increasingly replacing traditional electronic-structure calculations.
• A key question is whether to use pure machine learning (ML) or integrate it with physical principles.

Purpose of the Study:

• To explore an integrated modeling approach combining ML with physics for electronic structure calculations.
• To assess the benefits of intertwining data-driven techniques with physical approximations.

Main Methods:

• Developed a symmetry-adapted ML model of an effective Hamiltonian.
• Trained the ML model to reproduce electronic excitations from quantum-mechanical calculations.
• Utilized a parametrization corresponding to a minimal atom-centered basis.

Main Results:

• The integrated model accurately predicts properties for larger, more complex molecules than those used in training.
• Achieved significant computational savings by indirectly targeting calculation outputs.
• Demonstrated improved transferability and interpretability of ML models without compromising accuracy or efficiency.

Conclusions:

• Combining data-driven ML with physical approximations offers a powerful approach for materials modeling.
• This integrated strategy provides a blueprint for developing advanced ML-augmented electronic-structure methods.
• The approach enhances the reliability and applicability of ML in computational science.