Weisfeiler-Lehman subtree encoding for Bayesian optimization of atomic configurations

Abstract

The efficiency of Bayesian optimization (BO) of atomic configurations depends strongly on how configurations are encoded. We introduce the Weisfeiler-Lehman (WL) subtree kernel, which views configurations as element-labeled graphs and measures their similarity by how many local structural patterns they share, into Bayesian-optimization-based configuration search. Because this kernel is reproduced as the plain inner product of explicit features (L$^2$-normalized histograms of local topological patterns), introducing it reduces to introducing the corresponding features: the encoding enters existing BO frameworks as an ordinary descriptor. In a benchmark ground-state configuration search of cubic BC$_2$N evaluated with a universal machine-learning interatomic potential, the WL encoding reached the ground state almost immediately after a shared random initialization of 100 samples in every one of five independent rounds (108$\pm$5 evaluations on average), whereas the one-hot baseline required 280$\pm$122 evaluations; the WL-driven sampler first exhausted the degenerate ground-state group and then discovered the metastable degenerate groups from the bottom up, in order of increasing energy.