Dual-Level Training of Gaussian Processes with Physically Inspired Priors for Geometry Optimizations.

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Gaussian process (GP) regression is an emerging technique for molecular geometry optimization.
• The prior mean function is a critical component influencing GP performance.

Purpose of the Study:

• To design and validate physically inspired prior mean functions for GP regression in molecular geometry optimization.
• To implement and assess a dual-level training (DLT) optimizer for posterior-type priors.

Main Methods:

• Developed and tested force-field-based and posterior-type prior mean functions.
• Implemented a DLT optimizer incorporating classical mechanics and surrogate potential energy surfaces (PESs).
• Compared GP optimizer performance with constant prior against DLT and L-BFGS methods.

Main Results:

• Physically inspired priors (force-field and posterior-type) reduced optimization steps by 2-3x compared to constant-prior GP and L-BFGS.
• DLT optimizer demonstrated enhanced performance by leveraging physics-informed priors.
• Successfully demonstrated GP's potential in recovering real PESs using force-field priors.

Conclusions:

• Incorporating domain knowledge, specifically physics-based priors, significantly enhances GP models for molecular geometry optimization.
• The developed DLT optimizer offers a robust learning model for exploring molecular PESs and optimizing geometries.