Exploration versus Exploitation in Global Atomistic Structure Optimization

Abstract

The ability to navigate vast energy landscapes of molecules, clusters, and solids is a necessity for discovering novel compounds in computational chemistry and materials science. For high-dimensional systems, it is only computationally feasible to search a small portion of the landscape, and hence, the search strategy is of critical importance. Introducing Bayesian optimization concepts in an evolutionary algorithm framework, we quantify the concepts of exploration and exploitation in global minimum searches. The method allows us to control the balance between probing unknown regions of the landscape (exploration) and investigating further regions of the landscape known to have low-energy structures (exploitation). The search for global minima structures proves significantly faster with the optimal balance for three test systems (molecular compounds) and to a lesser extent also for a crystalline surface reconstruction. In addition, global search behaviors are analyzed to provide reasonable grounds for an optimal balance for different problems.

Accelerating crystal structure prediction by machine-learning interatomic potentials with active learning

“…Machine learning methods have been successfully used for crystal structure prediction problems. For example, an idea to fit a potential while structure search was suggested in 15 ; in 16 the authors apply active-learning techniques to predict the surface reconstructions; in 17 Bayesian optimization was used for the problem of prediction of molecular compounds.…”

Local Bayesian optimizer for atomic structures

“…It could potentially lead to a further reduction of the number of function evaluations [13]. The uncertainty provides a measure of how much a region of configuration space has been explored and can thereby guide the search also in global optimization problems [16,34,36].…”

“…Machine learning for PES modeling has recently attracted the attention of the materials modeling community [5][6][7][8][9][10][11][12][13][14][15][16][17][18]. In particular, several methods have focused on fitting the energies of electronic structure calculations to expressions of the form…”

Machine Learning for Computational Heterogeneous Catalysis

“…Traditional optimization methods (e. g. gradient‐based, Newtonian dynamics) often require calculating the energy and/or forces of the system at each optimization step. Therefore, machine learning (ML) surrogate models have been recently proposed as promising alternatives . The surrogate ML model should provide an approximation of the “true” first‐principle PES using only a few first principle calculations.…”

“…Therefore, machine learning (ML) surrogate models have been recently proposed as promising alternatives. [149][150][151][152][153] The surrogate ML model should provide an approximation of the "true" first-principle PES using only a few first principle calculations.…”