Online parameter estimation and model maintenance using parameter-aware physics-informed neural network

Machine learning-based (ML) digital twins for chemical processes are gaining popularity with the advent of Industry 4.0. These digital twins are often developed under the assumption of constant process parameters. However, in most chemical engineering processes, parameters often change during operations. To ensure optimal performance under such evolving conditions, there is a need for models that can adapt to these changes. In this work, we propose a framework for developing a PINN-based (Physics-Informed Neural Network) digital twin that is sensitive to parameter variations. The proposed framework also monitors the process in real-time using physics-based residual equations, identifies the parameters undergoing changes using sensitivity matrices, and re-estimates them to maintain the performance of the PINN model. We demonstrate the utility of the framework through a case study involving a continuous stirred tank reactor experiencing changes in activation energy and the overall heat transfer coefficient. The results show that the proposed framework improves the predictive accuracy of the PINN by approximately 84% for ramp changes and 12% for step changes in parameters. The framework is further applied to more realistic case studies, including a polymethyl methacrylate polymerization reactor and a pressure swing adsorption process, highlighting its applicability to high-dimensional nonlinear systems and cyclic separation processes. These findings indicate that the performance of digital twins can be significantly enhanced in the presence of varying process parameters by employing a PINN architecture that incorporates parameters as inputs and solves real-time inverse problems to estimate parameter values.