Multi-Fidelity Surrogate Modeling for Chemical Processes with Physics-Informed Neural Networks

In this study, a multi-fidelity surrogate modeling method based on physics-informed neural network (PINN) was proposed, which integrates high-fidelity simulation data and low-fidelity governing equations described by differential equations. By leveraging governing equations in the training of deep neural networks, the reliance on large amount of data has been relaxed. In the meantime, imposing physical laws ensures that the achieved surrogate models have clear physical meanings, which also improves the extrapolation performance of the models. Herein, the proposed multi-fidelity PINN surrogate modeling method was implemented to the simulation of the startup phase of a continuous stirred-tank reactor (CSTR) for illustrating its feasibility and advantages. From the computer experiment results, it is observed that the proposed method successfully reduced the sample size needed in model training and significantly improved the model extrapolation performance, facilitating its potential industrial applications.