Practical integration of machine learning into ab initio calculations and workflows: Accelerating the SCF cycle via...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Data-driven methods accelerate electronic structure calculations but often lack transferability.
• Training universal models requires extensive data and system-specific fine-tuning.

Purpose of the Study:

• To demonstrate a novel method for combining system-specific electronic structure machine learning (ESML) models.
• To improve computational efficiency in ab initio calculations by enhancing the initial guess for self-consistent field cycles.

Main Methods:

• Developed a
• stitched
• density matrix approach combining contributions from multiple ESML models.
• Applied the method to sequential calculations like geometry optimization and molecular dynamics.
• Integrated ESML with density matrix extrapolation algorithms.

Main Results:

• Successfully combined density matrices from system-specific ESML models.
• Achieved improved initial guesses for self-consistent field cycles, leading to computational speed-up.
• Demonstrated acceleration of standard computational calculations for water clusters and methane clathrate.

Conclusions:

• The synergistic use of ESML models and density matrix extrapolation significantly accelerates computational chemistry.
• This approach offers broad opportunities for hybrid quantum mechanical/machine learning (QM/ML) and ML/ML paradigms.
• Significant computational speed-ups are attainable for molecular and materials simulations.