A PINN Surrogate Modeling Methodology for Steady-State Integrated Thermofluid Systems Modeling

Physics-informed neural networks (PINNs) were developed to overcome the limitations associated with acquisition of large training data sets that are commonly encountered when using purely data-driven machine learning methods. This paper proposes a PINN surrogate modeling methodology for steady-state integrated thermofluid systems based on mass, energy, and momentum balance equations, combined relevant component characteristics fluid property relationships. The is applied two encapsulate important phenomena typically encountered, namely: (i) heat exchanger network different streams components linked in series parallel; (ii) recuperated closed Brayton cycle various turbomachines exchangers. results generated models compared benchmark solutions via conventional, physics-based process models. largest average relative errors 0.17% 0.93% cycle, respectively. It was shown use hybrid Adam-TNC optimizer requires between 180 690 fewer iterations during process, thus providing significant computational advantage over pure Adam optimization approach. resulting can make predictions 75 88 times faster than their respective conventional highlights potential as valuable engineering tool system design optimization, well real-time simulation anomaly detection, diagnosis, forecasting.