A new active learning approach for global optimization of atomic clusters

Multiscale molecular modelling: from electronic structure to dynamics of nanosystems and beyond

“…As for our own work, initial forays have been aimed at the global optimization of nanoparticle structures, with a focus on active learning methods, using small amounts of data to begin with and building the dataset progressively guided by Gaussian processes or neural networks. [153][154][155][156] Our plan is to extend this work in the next phase to finding transition states and then to full-fledged ML-accelerated dynamics for free-energy simulations. Although ML is often viewed as ''interpolation'' amongst sometimes huge datasets, our initial experience with the active learning protocols indicates some success at ''extrapolation'' combining local searches (exploitation) with more wide-ranging extrapolated exploration of potential energy surfaces.…”

Automatic structural elucidation of vacancies in materials by active learning

“…In this work, we present an Artificial Intelligence 20 (AI) method for non-stoichiometric structural elucidation for materials that have vacancies, based on a new active learning 21–23 (AL) protocol, where the vacancies are considered in the global structural search loop, using a structural descriptor. AL is based on supervised machine learning (ML), where the uncertainty is obtained from regression models.…”

“…Although the EI, PI and LCB acquisition functions have been evaluated from GP in this work—where the regression mean ( μ ) and its uncertainty ( σ ) are obtained analytically from Bayesian statistics, 30 they also can be calculated by frequentist methods 30 by using Support Vector Regressor 24,33,34 or Artificial Neural Network 21,22,35 regression algorithms with K -fold cross-validation 36 or bootstrap 37 for uncertainty quantification. 34 In addition, frequentist uncertainty estimates to obtain the EI can be done by random forests.…”

“…QMLMaterial is currently under development with the aim of providing a general framework for automatic structural elucidation coupled with first-principles methods for important molecular problems found in chemistry, physics and materials science. For instance, structural elucidation of nanoparticles 22 and the spin multiplicity, 23 doping in materials 21 and adsorption site search on surfaces. 35…”

“…23 The GA for defects is available in the Genetic Algorithm Software for Material Design & Discovery (GAMaterial). 22…”