Adaptive Pruning for Increased Robustness and Reduced Computational Overhead in Gaussian Process Accelerated Saddle...

Area of Science:

• Computational Chemistry
• Chemical Physics
• Machine Learning in Chemistry

Background:

• Gaussian process (GP) regression is a powerful tool for accelerating computationally expensive calculations in chemistry, such as saddle point searches.
• Traditional GP regression methods face challenges with computational overhead during hyperparameter optimization and potential failures in poorly represented regions of the potential energy surface.
• Efficiently navigating high-dimensional energy surfaces is crucial for understanding chemical reactions and molecular properties.

Purpose of the Study:

• To address the inefficiencies and failures associated with GP regression in accelerating saddle point searches.
• To develop a more robust and scalable GP-based algorithm for exploring high-dimensional potential energy surfaces.
• To reduce the computational cost associated with evaluating energies and forces in complex chemical systems.

Main Methods:

• Implemented geometry-aware optimal transport measures and an active pruning strategy using Wasserstein-1 distances for farthest-point sampling.
• Selected geometrically diverse configurations to manage the increasing cost of GP updates.
• Introduced a permutation-invariant metric for stability and a logarithmic barrier penalty to control signal variance growth.

Main Results:

• Successfully reduced the mean computational time by over 50% for a dataset of 238 challenging chemical reaction configurations.
• Demonstrated enhanced stability and robustness of the GP approach, even in complex, high-dimensional systems.
• Validated the efficacy of the physically motivated algorithmic improvements in practical applications.

Conclusions:

• The enhanced GP regression algorithm offers a robust and scalable solution for accelerating saddle point searches in computational chemistry.
• These improvements significantly decrease computational demands, making GP regression a more practical tool for analyzing chemical reactions.
• The developed methods provide a reliable strategy for navigating complex energy landscapes where energy and force evaluations are costly.