Efficient Implementation of Gaussian Process Regression Accelerated Saddle Point Searches with Application to...

Area of Science:

• Computational Chemistry
• Chemical Physics
• Materials Science

Background:

• Locating saddle points is crucial for understanding reaction mechanisms and rates in thermally activated processes.
• High-dimensional energy surfaces pose computational challenges for electronic structure calculations.
• Transition state theory relies on accurate saddle point identification for rate estimations.

Purpose of the Study:

• To develop an efficient Gaussian process regression (GPR) accelerated method for locating saddle points.
• To reduce the number of electronic structure calculations required for convergence.
• To assess the performance of the GPR method on a diverse set of molecular reactions.

Main Methods:

• Implementation of Gaussian process regression (GPR) to accelerate the minimum mode following method.
• Utilizing a dimer approach to estimate the lowest eigenmode of the Hessian.
• Constructing and updating a surrogate energy surface with each electronic structure calculation.

Main Results:

• Achieved an order of magnitude reduction in electronic structure calculations needed for saddle point identification.
• Demonstrated comparable efficiency to internal coordinate methods (e.g., Sella) using Cartesian coordinates.
• Showed reduced wall times for saddle point searches in most cases, even at a low Hartree-Fock level.

Conclusions:

• GPR-accelerated saddle point searches offer significant computational savings.
• The method is robust across molecular systems with varying degrees of freedom stiffness.
• Efficient GPR implementation enables faster exploration of reaction pathways in computational chemistry.