Fig. 1. Schematic classification of popular molecular machine learning potentials (MLPs). Representative models on the scheme (models tested here in bold): (1) kernel methods with global descriptor (KM-GD: KREG,18 KRR-CM,41 and sGDML20), (2) kernel method with fixed local descriptor (KM-fLD: GAP42-SOAP,43 KRR-FCHL,44 and KRR-aSLATM45), (3) neural networks with fixed local descriptor (NN-fLD: ANI,46 DPMD,47 and BPNN48), (4) neural networks with learned local descriptor (NN-lLD: PhysNet,36 SchNet,49,50 and MEGNet51).
