One Bayesian loop for every energy-surface search: minima, saddles, and reaction paths

What is it about?

Almost every question in computational chemistry comes down to searching a potential energy surface: where a molecule sits at rest, which barrier it crosses to react, what path it takes between two states. Each step of that search asks an electronic structure code for the energy and forces at a new geometry, and each of those answers can cost minutes to hours of supercomputer time. The standard way to cut the bill is to fit a cheap surrogate model on the points already computed and let it propose where to look next. This invited tutorial review shows that the three searches usually taught as separate subjects are one and the same algorithm: minimization, single-point saddle searches (the dimer or minimum-mode-following method), and double-ended reaction-path searches (the nudged elastic band). All three run a six-step surrogate loop and differ only in the inner target they optimize and the rule they use to pick the next point. The surrogate is a Gaussian process trained on energies and their gradients, built on inverse-distance kernels, and driven by active learning. Done correctly, it reduces the number of expensive evaluations by roughly an order of magnitude. The review walks through the math from first principles and comes with pedagogical Rust code that runs all three searches from the same loop.

Featured Image

Photo by Joshua Dixon on Unsplash

Photo by Joshua Dixon on Unsplash

Why is it important?

Surrogate-accelerated saddle and path searches are scattered across two decades of papers, each with its own notation, kernel choice, and convergence tricks. A newcomer faces a wall. By collapsing minimization, saddle searches, and path searches into one Bayesian optimization loop, the review turns that wall into a single entry point: learn the loop once, reuse the kernel, hyperparameter, and trust-region machinery everywhere. The review also states plainly when surrogates pay off and when they do not. The order-of-magnitude saving depends on how expensive the oracle is, how far the search has to travel, and whether analytical forces are available; the review states those conditions instead of advertising a universal speedup. Four production extensions move a textbook method into a working high-throughput pipeline: farthest-point sampling with an Earth Mover's Distance metric, MAP regularization, an adaptive trust radius, and random Fourier features for scaling.

Perspectives