Adaptive physics-informed neural network-based digital twins integrated with Ensemble Kalman Filter

With the advent of Industry 4.0, the growing use of machine learning-based predictive models has accelerated their adoption as digital twins for online applications. However, most existing models assume constant process parameters and lack real-time updating mechanisms. In reality, process parameters change due to various factors, and the parameters in physics-based models may also be inaccurate. This highlights the need for a parameter-sensitive digital twin that can adapt to such changes. In this work, we introduce a novel, robust, adaptive digital twin based on a physics-informed neural network (PINN), capable of adjusting in real time to changes in process parameters to maintain the performance of predictive models. Model updates are performed on the deep neural network using the Ensemble Kalman Filter (EnKF). We implement the adaptive digital twin on a real-world polymerization process for Polymethyl-Methacrylate (PMMA), considering various scenarios involving process parameter changes, model-plant mismatches, and different measurement scenarios. The results demonstrate a 10- to 20-fold improvement in accuracy compared to digital twins that do not incorporate parameter updates. The adaptive PINN also achieves approximately a 96% reduction in computation times compared to the EnKF with a physics-based model predictor, making it highly suitable for online applications. This emphasizes that the physics-based model is only necessary for training the PINN offline, can be completely discarded during online implementation. This approach provides a recursive, hybrid modeling framework that is modularizable and scalable, generalizable, robust, and maintains model performance under changing environmental conditions. • We develop an adaptive PINN digital twin that is sensitive to process parameters. • State and parameter estimation enables model maintenance and adaptation of the PINN. • The digital twin performs comparably to a first principles twin and scales better.