Fig. 2. Schematic of the Δ-learning concept and application. A three-dimensional (3D) molecular conformation is used as an input to either a single- or multi-task trained message-passing neural network (MPNN). When a Δ-learning endpoint is requested, an additional GFN2-xTB calculation is carried out, and the network is tasked with predicting the correction (Δ) between this baseline value and its ωB97X-D/def2-SVP analogue. If a direct-learning prediction is requested, the network outputs an approximation to the ωB97X-D/def2-SVP value, without using the GFN2-xTB baseline.
