Fig. 3. Learning curves of Δ- and direct-learning model for extensive (panels A and H) and intensive (panels B–E) molecular properties, and for atom- and bond-centered properties (panels F and G). These show the mean absolute error (MAE) as a function of training set size (axes in logarithmic scale) computed on the three test sets comprised of ∼29 k molecules (∼88 k conformers). For the energy models (panels A and H), all conformers of the same molecule are grouped within the same train-validation-test splits (see Methods). For intensive molecular properties (panels B–E), learning curves are shown for both single-task and multi-task learning paradigms. Chemical accuracy thresholds are indicated for energy and orbital energy models (1 kcal mol⁻¹ ≈ 43.4 meV).³⁹.