Control-Oriented Modeling for Industrial Propylene Polymerization Process based on Physics-Informed Neural Network

Loop reactors are extensively employed in the industrial production of polypropylene. Nonetheless, the high nonlinearity and stiffness of this process pose significant challenges. Traditional methods are time-consuming in simulating this process. This work is aimed at using Physics-Informed Neural Network (PINN) to improve the computation efficiency. PINN can use neural network to directly give the solution without solving the ODEs through explicit numerical methods. Particularly, we extend the application of PINNs to include initial states and control inputs, making them suitable for control tasks. In order to enable the model to track the stiffness variables and to improve the generalization performance, we combine Latin Hypercube Sampling (LHS) Gaussian-Lagrange method for sampling time, states and control collocation points. Furthermore, we utilize the Extended Physics-Informed Neural Networks framework, which ensures that solutions inherently satisfy initial conditions and constraints. The Control-Oriented Physics-Informed Neural Network (COPINN) proposed here achieves calculation accuracy comparable to the Backward Differentiation Formula (BDF) method while significantly improving computational efficiency.