Fig. 1: Workflow in the Plan for Robust and Accurate Potentials (PRAPs) package.
From: Machine learned interatomic potentials for ternary carbides trained on the AFLOW database
PRAPs automates the generation of a moment tensor potential (MTP), given a set of data generated from quantum-mechanical calculations (here the online AFLOW database32), and/or a pre-generated set of structures (here RANDSPG70). Five MTPs are trained on a set of configurations (cfgs) and the best (the pre-Robust Potential, pre-RP) is chosen (gray box). The pre-RP potential and training set are employed to initialize the training of the RP via an active learning scheme (ALS, blue box). This requires iterations of relaxing and retraining on structures in, and chosen from, the Relaxation Set. After training, the RP relaxes everything in the Relaxation Set, and the results are filtered to retain only the lowest-energy configurations in each composition (Low-Energy Robust Relaxed). This is then used to train the Accurate Potential (AP) via active learning (orange box). The AP-relaxed structures may be sent for relaxation via AFLOW-DFT, followed by subsequent analysis (green box). The inset schematically illustrates the energy distribution of the structures that can be relaxed with the RP and the AP. Rectangles with sharp corners represent structural data, while curved corners represent potentials.
