Learning to Optimize Molecular Geometries Using Reinforcement Learning.

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Chemistry

Background:

• Quasi-Newton methods, including BFGS, are standard for molecular geometry optimization.
• These methods can struggle with initial geometries far from the energy minimum due to inaccurate quadratic approximations.

Purpose of the Study:

• To develop a reinforcement learning model that improves quasi-Newton optimization performance.
• To create a correction term for the BFGS algorithm to handle difficult starting points.

Main Methods:

• A reinforcement learning model was trained to predict a correction term for the BFGS update step.
• The model was tested on molecular geometry optimizations starting from perturbed initial guesses.

Main Results:

• The reinforcement learning-enhanced BFGS method completed optimizations in approximately 30% fewer steps compared to pure BFGS for perturbed geometries.
• The new approach demonstrated comparable convergence to BFGS with line search but was significantly faster by reducing gradient evaluations.

Conclusions:

• Reinforcement learning offers a promising strategy to augment traditional optimization algorithms in computational chemistry.
• The proposed method accelerates molecular geometry optimization, particularly for challenging initial configurations, without sacrificing convergence accuracy.