Learning the One-Electron Reduced Density Matrix at SCF Convergence Thresholds.

Area of Science:

• Computational chemistry
• Materials science
• Quantum mechanics

Background:

• Conventional electronic structure methods are computationally expensive.
• Machine learning (ML) offers a potential for computationally efficient surrogates.
• Accurate prediction of the one-electron reduced density matrix (1-RDM) is crucial for electronic structure calculations.

Purpose of the Study:

• To develop ML models that accurately predict the 1-RDM from electron-nuclear potentials.
• To reduce the training data requirements for accurate 1-RDM prediction.
• To enable stable ab initio molecular dynamics using ML-predicted 1-RDMs.

Main Methods:

• Training ML models to map electron-nuclear interaction potentials to the 1-RDM.
• Implementing targeted model optimization strategies to reduce training set size.
• Developing a force-correction algorithm for ML-powered ab initio molecular dynamics.

Main Results:

• ML models achieve 1-RDM prediction accuracy within a standard self-consistent field (SCF) threshold.
• Substantially smaller training set sizes are sufficient compared to previous work.
• Stable ab initio molecular dynamics simulations are enabled for molecules up to biphenyl size.

Conclusions:

• ML-based 1-RDM prediction is a viable and efficient alternative to conventional methods.
• Optimized ML models significantly reduce data requirements for high accuracy.
• The developed force-correction algorithm extends the applicability of ML surrogates to larger molecular systems and dynamics simulations.