Abstract
Variable-order fractional partial differential equations describe heterogeneous anomalous diffusion but can be costly to solve repeatedly at scale because of nonlocal memory and irregular geometry. We propose a physics-informed variable-order graph neural operator (PI-VO-GNO) in which the neural architecture is identical to a data-only graph operator and fractional calculus enters only through a differentiable variable-order Caputo residual. Across three paired seeds, the correct variable-order physics reduces prediction error by 10.44% in distribution and 5.17–6.90% under three distribution shifts, while reducing the true-order discrete residual by 37.7–46.8%. With only 10% of the training trajectories, prediction gains increase to 27.2–29.1%. Under 10% Gaussian corruption of the 25% training subset, PI-VO-GNO retains 13.2–18.6% lower prediction error. A model trained on 128-node graphs transfers to 192- and 256-node discretizations with 8.85–14.54% lower error. The physics loss adds no trainable parameters, only 7.7% training-step overhead, and negligible inference overhead. Reference-solver refinement exhibits monotonic spatial and temporal self-convergence. Mismatched integer- and fixed-order physics instead degrade prediction. Under a severe finite-element cross-discretization shift, field-error gains become nearly neutral, although physical-consistency gains persist. An interpolation control yields 2.70% mean trajectory discrepancy between inverse-distance and piecewise-linear (P1) transfer.