(543b) A Data Reconciliation Methodology for Reduced Order Modeling of Process Systems

Recent AI successes in computer vision, game playing, and search problems, have contributed to renewed interest in the usage of such techniques in science and engineering. Unlike these applications, complex dynamical systems in science and engineering are governed by fundamental laws, and constitutive equations. Black-box machine learning applications do not appropriately take these into accounts a priori, and are thus unable to reliably learn the underlying physics of such systems. With the gaining popularity of physics-informed neural network approaches, strides are being made in the hybrid AI domain. However, such an approach forgoes interpretability due to the nature of neural network models. On the other hand, reduced-order modeling approaches can yield promising results as the complexity of process systems increases. Here, we present a symbolic reduced-order modeling approach that ensures adherence to fundamental constraints. We demonstrate the benefits of the same on noisy data, for multiple unit operations and processes. Such a modeling methodology permits enhanced explainability, in comparison to conventional AI models.