Figure2

Figure 2. Schematic of the different multi-fidelity methods. (A) Transfer Learning, a model is sequentially trained on $$ {\rm{PBE}} \cup{\rm{HSE}} \cup{\rm{EXP}} $$, $$ {\rm{HSE}} \cup {\rm{EXP}} $$, and $$ {\rm{EXP}} $$; (B) Joint Learning, a model learns on multiple targets at once in parallel, by using a shared architecture; (C) Stacking Ensemble Learning, three submodels are separately trained on the three data sources and then a linear regression (LR) is fitted from the submodel predictions to the experimental value; and (D) Deep-Stacking Ensemble Learning, the same principle as stacking, except the last layer of each submodel is fed to a neural network.