(149y) Physics-Informed Neural Networks (PINNs) for Process Systems with Model Plant Mismatch

This work addresses the problem of incorporating process knowledge into machine learning methods and illustrates it for a stirred-tank reactor system. In particular, purely data-driven modeling techniques like subspace identification and feed-forward neural networks (NN) use data analysis to predict relationships between system state variables. These predictions may violate the physics of the system. By using a hybrid modeling technique like a physics-informed neural network (PINN), first-principles knowledge is incorporated into data-driven modeling, resulting in a model that respects known physical relationships. The PINN’s implementation is demonstrated as follows: data is obtained from a 2 CSTRs in series simulation system. A standard NN and a regular subspace model are developed first. While both the regular subspace and NN models capture the process dynamics reasonably well, the resultant models predict (small but non-negligible) changes in outlet concentrations from CSTR 1 when changes are made to the inlets to CSTR 2. Subsequently, a PINN is developed based on the known system of differential equations (using a model with parametric mismatch to demonstrate robustness). The PINN not only captures the process dynamics reasonably well, but is also consistent with the physics of the problem. As such, it demonstrates superior results compared to the regular NN and subspace models. These results indicate the benefit of utilizing hybrid modeling techniques.