Gaussian process-based active learning
Here, we show that active learning based on Bayesian inference can accelerate and automate the training of high-quality force fields by making use of accurate internal estimates of model error. By combining DFT with Gaussian process regression during molecular dynamics simulations, accurate force fields for a range of single- and multi-element systems are obtained with ~100 DFT calls. Moreover, we demonstrate that the model can be flexibly and automatically updated when the system deviates from previous training data. Such a reduction in the computational cost of training and updating potentials promises to extend ML modeling to a much wider class of materials than has been possible to date. The method is shown to successfully model rapid crystal melts and rare diffusive events, and so we call our method FLARE: Fast Learning of Atomistic Rare Events, and make the open-source software freely available online.
For more information, check out our preprint! ArXiv preprint
The code is available and fully documented, check it out!