摘要
•A PIHF surrogate integrates CFD data with low-fidelity physical constraints.•Low-fidelity mass conservation is embedded into the DNN training loss.•PIHF achieves lower extrapolation RMSE than FFNN, LF, and MFGP models.•PIHF exhibits reduced performance variability across different datasets.•PIHF enables process optimization within 37–147 s instead of hours-long CFD runs.
Deep neural networks are increasingly used as surrogate models for chemical process analysis and optimization. However, their application is often limited by the high computational cost and long simulation time required to generate large high-fidelity training datasets, restricting their reliability in extrapolation and real-time optimization.
To address this limitation, a physics-informed hybrid-fidelity (PIHF) surrogate modeling framework is proposed. High-fidelity transient datasets are generated using three-dimensional computational fluid dynamics simulations in ANSYS Fluent for the start-up stage of a non-isothermal continuously stirred tank reactor under varying inlet concentration, volumetric flow rate, and inlet temperature. Low-fidelity physical knowledge derived from mass conservation and simplified reactor assumptions is embedded into the training loss of a deep neural network, enabling physically consistent learning under small-sample conditions.
Case studies show that PIHF achieves an average extrapolation RMSE of 0.182, outperforming a feedforward neural network (0.391), a low-fidelity simulation model (0.468), and a multi-fidelity Gaussian process model (0.298). In inverse optimization tasks, PIHF reduces computation time to 37–147 seconds, compared to over 11,000 hours for high-fidelity simulations, demonstrating its accuracy and efficiency for complex chemical process applications.
[Display omitted]