Physics-Constrained Deep Learning for Isothermal CSTR

This research study investigates the approach of using physics-constrained deep learning in modelling isothermal continuous stirred-tank reactor (CSTR) to address the challenges in its current process control and optimisation. An inaccurate system identification affects prediction and consequently deteriorates the control performance. Physics-constrained deep learning is a promising machine learning framework that can better govern the system dynamics. Therefore, this research study attempts to investigate its application in predicting the behaviour of isothermal continuous stirred-tank reactor, particularly in modelling the concentration of reactant at the outlet of the reactor. The research methodology comprises data preparation, network architecture design, model training, model validation, and solution prediction. Different activation functions, optimizers, and epochs are used in the design. The prediction made by physics-constrained deep learning converged to that of the exact solution whereby the lowest error obtained at 4000 epochs is 2.1076e−5, when using Adam optimizer and tanh activator in the design. Increasing the number of epochs increases the prediction accuracy. The selection of the network architecture requires extensive numerical experimentation and is often depending on the problem.