Fig. 1: Our MLIP training strategy. | npj Computational Materials

Fig. 1: Our MLIP training strategy.

From: Machine-learning potentials for nanoscale simulations of tensile deformation and fracture in ceramics

Fig. 1

Schematic visualization of a our general training procedure (Procedure 1, Section Training procedure and fitting initial MLIPs), and b up-fitting (Section MLIPs’validation against atomic scale tensile tests and MLIPs’ up-fitting for nanoscale tensile tests), both applicable to any MLIP formalism with the extrapolation/Maxvol grade (γ) quantification, particularly MTP (used in this work) and ACE. c An overview of the here-developed MLIPs: MLIP-[0001], MLIP-[10\(\overline{1}0\)], MLIP-[\(\overline{1}2\overline{1}0\)] (applicable for atomic scale tensile loading, Sections Training procedure and fitting initial MLIPs and MLIPs’validation against atomic scale tensile tests); MLIP-[1], MLIP-[2], MLIP-[3] (intermediate MLIPs produced by up-fitting, Section MLIPs’ up-fitting for nanoscale tensile tests); and MLIP-[4] (applicable for both atomic and nanoscale tensile loading, Section MLIPs’ up-fitting for nanoscale tensile tests--Otherloading conditions and MLIP’s transferability). Details of the training datasets are given in the corresponding sections and additionally summarized in the Methods under “Development of machine-learning interatomic potentials”. both.

Back to article page