Heuristic Global Optimization in Chemical Compound Space.

Abstract

We report developments in combinatorial optimization under constraints in chemical space. We considered random functions, which serve as a baseline to measure performance, and constrained optimizations over two databases of electrochromic molecules (∼5000 and 1012, respectively). These problems were optimized using sequential heuristic next-neighbor search (HNNS) (introduced in Elward and Rinderspacher Phys. Chem. Chem. Phys. 2015, 17, 24322 and Elward and Rinderspacher Mol. Syst. Des. Eng. 2018, 3, 485) and kernel-based efficient global optimization (EGO) with two reordering strategies. In addition to the average ordering method introduced with sequential HNNS, a new reordering based on a locally separable linear model is formulated and applied. Presented is the analysis of periodic kernels for EGO: the modified Dirichlet kernels D̃n(x) = ∑in[cos(2πix) - (-1)i]/i, the minimalist kernel KN(x) = cos(πx/N)2/2 - δx/2, and an analogue to the popular Gaussian kernel, GNσ(x) = e-|xmodN|2/(2σ2), where n is the order of the Dirichlet kernel, N is the number of choices on a combinatorial site, and σ is a broadening factor. We find that reordering is pivotal for high optimization efficiency, in both the average and worst-case scenarios. The new linear-estimation ordering paradigm is superior in the chemistry context. Furthermore, with judicious use of hyperparameters and algorithmic choices, EGO outperforms HNNS. The global optimum for the small chemical problem is found with >98% likelihood for all global search methods employing a linear reordering heuristic for organizing the search space.