A dual-graph deep learning network that hard-codes petroleum physics alongside learned statistics cuts oil production prediction error to about 20 percent, far outperforming conventional models.