A Physics-Informed Neural Network Surrogate for Two-Dimensional Field Prediction in a Steam Methane Reforming Reactor

Steam methane reforming (SMR) is an important industrial reaction. Process simulation tools such as Aspen and computational fluid dynamics (CFD) are widely used to model it and perform well under specific operating conditions. However, when operating conditions vary and repeated calculations are required, the computational time cost becomes high. Neural networks can make fast data-driven predictions, but purely data-driven models may produce results that violate physical laws. Physics-informed neural networks (PINNs) combine data-driven prediction with physical equations, using the governing partial differential equations as constraints during training to improve physical consistency. In this work, a dual-network PINN was trained on 400 of 520 numerical reference conditions, with 60 conditions used for validation and 60 for testing. The trained PINN achieves mean casewise RMSEs of 3.08 K for bed temperature and 0.0053 bar for pressure, with a recorded reference-to-PINN evaluation-time ratio of approximately 1.81×104. Within the investigated operating range, RBF interpolation and ordinary Kriging achieved lower prediction and balance errors and shorter prediction times than the PINN, whose inlet-region temperature accuracy remains a limitation.