GP-accelerated saddle point searches with 10x fewer force evaluations

What is it about?

Finding transition states -- the atomic configurations at energy barriers between reactant and product -- requires many evaluations of quantum mechanical energies and forces, each of which can take hours. Gaussian process (GP) regression can reduce the number of these evaluations by building a surrogate energy surface on the fly, but previous implementations were slow to set up and hard to use. We reimplemented the GP-accelerated dimer method for the EON saddle point search code, focusing on practical performance. The implementation works in simple Cartesian coordinates rather than requiring specialized internal coordinates, yet achieves comparable accuracy. On a benchmark of 238 molecular reactions, the GP approach needs roughly one tenth the force evaluations of the standard dimer method. The code is publicly available in the EON software package (https://eondocs.org) and interoperates with standard electronic structure codes.

Featured Image

Photo by New Material on Unsplash

Photo by New Material on Unsplash

Why is it important?

Saddle point searches are a bottleneck in computational studies of reaction mechanisms, catalysis, and diffusion. Each search can require hundreds of force evaluations at the DFT or coupled-cluster level. Reducing that count by an order of magnitude makes it feasible to survey reaction networks that were previously too expensive. The Cartesian-coordinate formulation matters because it avoids the coordinate-transformation machinery that made earlier GP implementations fragile for complex systems. This simplicity also makes the method easier to couple with new electronic structure codes or machine-learned potentials. The benchmark on 238 diverse reactions provides a quantitative baseline for future methods.

Perspectives