What is it about?
Locating a first-order saddle is the expensive step of a harmonic transition-state-theory calculation. When each energy and force comes from an electronic-structure code, the number of those calls is the bill. A Gaussian process can build a local surrogate from the calls already made and propose the next geometry on that cheaper surface. Earlier implementations were slow to set up and often required internal coordinates. This is a C++ GP-dimer in eOn that stays in Cartesian coordinates. On the 500 starting configurations assembled by Hermes and coworkers (265 distinct molecules, run as singlets or doublets), the GP-dimer cuts electronic-structure evaluations by about an order of magnitude relative to a plain dimer. The same Cartesian search matches Sella's force-call count on that set without ghost atoms for near-linear arrangements. The code is in eOn (https://eondocs.org).
Featured Image
Photo by New Material on Unsplash
Why is it important?
A transferable machine-learned potential is the wrong tool for a single saddle search: the training set almost never samples the barrier. A per-search GP uses only the points that search generates. Cartesians matter because catalysis on a surface produces near-linear atom triples, and internal coordinates then need extra ghost atoms. The 500-configuration set is the one Sella already converges, so the comparison is on Sella's own ground. The later adaptive-pruning paper (238 harder cases, wall-time bound) sits on top of this implementation.
Perspectives
I kept needing a dimer that would survive a 500-configuration benchmark, not a handful of textbook saddles. The GP-on-a-dimer idea was already in the literature. The work was making it run in Cartesians inside eOn and still cut the force-call count by about ten. SURFsara, the Dutch national computing center, helped us benchmark at scale. The production runs were on Elja. Internal coordinates looked cleaner on paper. On systems with near-linear triples they introduced more problems than they solved, so I stayed in Cartesians. Matching Sella's force-call count on its own set, without ghost atoms, is the result I needed before trusting the method on a surface. This implementation is the baseline the later OT-GP pruning paper then sped up.
Rohit Goswami
University of Iceland
Read the Original
This page is a summary of: Efficient Implementation of Gaussian Process Regression Accelerated Saddle Point Searches with Application to Molecular Reactions, Journal of Chemical Theory and Computation, August 2025, American Chemical Society (ACS),
DOI: 10.1021/acs.jctc.5c00866.
You can read the full text:
Resources
Contributors
The following have contributed to this page
