Fig. 4: Benchmark of convergence of liquid-water properties for CAGO-based and standard active learning starting out from a single structure.
From: Learning atomic forces from uncertainty-calibrated adversarial attacks
We report data for the Allegro MLIP with target error 100 meV/Å and for DeePMD with target error 100 meV/Å, 200 meV/Å, as well as picking maximum uncertainty structure from MD and random structure sampling from 500 K MD simulations (see Methods IV F for more details). Each point in every graph has been computed from 12 models using the subset of stable MLIPs, with three replica MD simulations lasting 2 ns each, for a total of 540 Allegro and 3744 DeepMD simulations. The error bars are the standard deviation between the different committee members, and the horizontal lines correspond to the average final value of the property for Allegro CAGO. a The liquid water box structure that was used to start the active learning training. Oxygen is represented in red, with hydrogen in white. b Percentage of stable MLIPs in accordance with the stability criteria in Methods IV F. c Average force error of MLIPs on the liquid water structures. d Mean liquid water density. e Self-diffusion coefficient. f First coordination number (integral of the first radial distribution function peak).
