Expectation-Maximization-Based Optimization of Neural Quantum States for Ab Initio Quantum Chemistry.

Area of Science:

• Quantum Chemistry
• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Accurate ab initio quantum chemistry is vital for atomic-scale understanding.
• Neural quantum states (NQSs) offer improved accuracy but face high computational costs.

Purpose of the Study:

• To develop a computationally efficient method for optimizing NQS.
• To accelerate NQS optimization without sacrificing accuracy.

Main Methods:

• Proposed expectation-maximization-based optimization of NQS (EMO-NQS).
• Reformulated time-intensive operations into E-step and M-step for cost amortization.
• Integrated an active learning (AL)-based sampling method to enhance robustness.

Main Results:

• EMO-NQS achieved at least twice the optimization speed of the vanilla NQS method.
• Accuracy was maintained compared to traditional NQS methods.
• AL-based EMO-NQS demonstrated consistent acceleration and robust accuracy for molecular potential energy surface construction.

Conclusions:

• EMO-NQS effectively reduces computational costs for NQS optimization.
• The method provides a viable approach for accelerating quantum chemistry simulations.
• Active learning integration enhances the reliability of the EMO-NQS method.