Incorporating Domain Differential Equations into Graph Convolutional Networks to Lower Generalization Discrepancy