Enhancing classical gold nanoparticle simulations with electronic corrections and machine learning.

Classical simulations of materials and nanoparticles have the advantage of speed and scalability but at the cost of precision and electronic properties, while electronic structure simulations have the advantage of accuracy and transferability but are typically limited to small and simple systems due to the increased computational complexity. Machine learning can be used to bridge this gap by providing correction terms that deliver electronic structure results based on classical simulations, to retain the best of both worlds. In this study we train an artificial neural network (ANN) as a general ansatz to predict a correction of the total energy of arbitrary gold nanoparticles based on general (material agnostic) features, and a limited set of structures simulated with an embedded atom potential and the self-consistent charge density functional tight binding method. We find that an accurate model with an overall precision of 14 eV or 8.6% can be found using a diverse range of particles and a large number of manually generated features which were then reduced using automatic data-driven approach to reduce evaluation bias. We found the ANN reduces to a linear relationship if a suitable subset of important features are identified prior to training, and that the prediction can be improved by classifying the nanoparticles into kinetically limited and thermodynamically limited subsets based prior to training the ANN corrections. The results demonstrate the potential for machine learning to enhance classical molecular dynamics simulations without adding significant computational complexity, and provides methodology that could be used to predict other electronic properties which cannot be calculated solely using classical simulations.