Self-Consistency Error Correction for Accurate Machine Learning Potentials from Variational Monte Carlo

Area of Science:

• Computational materials science
• Quantum chemistry
• Machine learning

Background:

• Variational Monte Carlo (VMC) is a powerful method for training machine learning interatomic potentials (MLIPs).
• VMC training sets often use partially optimized wave functions (WFs) to reduce computational cost.
• Frozen variational parameters in WFs introduce a self-consistency error (SCE), biasing forces and pressures.

Purpose of the Study:

• To demonstrate the detrimental impact of SCE on MLIP accuracy.
• To apply a novel SCE correction method to VMC training data.
• To improve the reliability of MLIPs for molecular dynamics (MD) simulations.

Main Methods:

• Utilizing VMC to generate training data for MLIPs.
• Implementing an SCE correction for VMC wave functions with frozen Kohn-Sham orbitals.
• Training MLIPs on both uncorrected and SCE-corrected VMC data.
• Performing MD simulations to evaluate MLIP performance and physical observables.

Main Results:

• The self-consistency error (SCE) was shown to negatively impact MLIP accuracy, using high-pressure hydrogen as a test case.
• Applying the SCE correction to VMC training sets significantly improved MLIP quality.
• MLIPs trained on SCE-corrected data approached the accuracy of those trained on fully optimized WFs.
• MD simulations confirmed that SCE-corrected MLIPs yield more reliable physical observables.

Conclusions:

• The developed framework effectively corrects the self-consistency error in VMC training data.
• This correction enables the generation of high-quality MLIPs suitable for accurate MD simulations.
• The approach facilitates the creation of larger, more reliable VMC-based training datasets.