Physics-Informed Neural Networks for Process Systems: Handling Plant-Model Mismatch

This work addresses the problem of leveraging first-principles knowledge with data-driven techniques in the Physics-Informed/Inspired Neural Network (PINN) framework to handle plant–model mismatch. To this end, a PINN is developed utilizing the first-principles model of the system and plant data and demonstrated to handle plant–model mismatch. The PINN is compared with another dynamic modeling technique, a Recurrent Neural Network (RNN), and for the illustrative simulation example, is shown to improve the predictive capabilities of the model compared to the other techniques. In particular, purely data-driven approaches often encounter challenges when applied to complex systems. This can lead to compromised predictive performance in situations where the model fails to capture the actual relationships among system variables. In contrast, the PINN respects the physical characteristics of the problem, while yielding a good dynamic model, based on process data. These results indicate the benefit of utilizing hybrid modeling techniques and their potential application to more complex systems.