Exploring the Design Space of Machine Learning Models for Quantum Chemistry with a Fully Differentiable Framework.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Traditional atomistic machine learning (ML) models predict quantum mechanical (QM) properties directly.
• Emerging ML approaches predict intermediate QM calculation components like the electronic Hamiltonian.
• This allows deriving multiple properties via physics-based operations on ML predictions.

Purpose of the Study:

• To present a framework integrating ML Hamiltonian prediction with differentiable QM workflows.
• To enable indirect training of ML models against various QM properties.
• To explore the design space of hybrid ML/QM models and optimize their performance.

Main Methods:

• Developed a framework integrating effective electronic Hamiltonian prediction with PySCFAD, a differentiable QM workflow.
• Facilitated indirect model training against functions of the Hamiltonian (e.g., energy levels, dipole moments).
• Explored hybrid ML/QM model design choices, including multi-target training and reduced-basis Hamiltonians.

Main Results:

• Demonstrated the framework's ability to learn reduced-basis ML Hamiltonians reproducing targets computed on larger bases.
• Evaluated the accuracy and transferability of hybrid ML/QM models.
• Compared hybrid model performance against traditional ML surrogate models for atomic properties.

Conclusions:

• The integrated framework offers a flexible approach for designing and training hybrid ML/QM models.
• Incorporating multiple targets and reduced-basis Hamiltonians influences model performance.
• Findings guide the optimization of ML-QM interfaces for enhanced accuracy and transferability.