Fig. 3: Main steps of the active learning procedure.
From: Atomistic fracture in bcc iron revealed by active learning of Gaussian approximation potential
a Workflow of the active learning algorithm. (I) Fracture simulation by K-test at T = 0 K. (II) Evaluation of the GAP predicted uncertainty for each frame generated during K-test. (III) Detection of extrapolation by comparing the maximum model uncertainty of each configuration with the predefined criterion. (IV) Construction of the crack-tip cell that is computable by DFT. (V) DFT calculation of the crack-tip cell. (VI) Refitting GAP. Steps (I)–(VI) are repeated until convergence is achieved. b Construction of the crack-tip DFT cell. (I) Identification of the atom with the largest uncertainty (ID0) in the deformed configuration. (II) Selection of a group of atoms (IDg) in the square centred on ID0 in the undeformed state. (III) Identification of the crack tip in group IDg in the deformed state. (IV) Duplication and rotation of the selected region. (V) Generation of DFT cell containing two crack tips by merging the rotated replica and the original one.
