Bayesian LSTM framework for the surrogate modeling of process engineering systems

Abstract

This work explores the application of Bayesian Long Short-Term Memory (LSTM) networks as surrogate models for process engineering systems. We illustrate the model’s ability to adapt to parameter variations, making it a robust asset in optimizing real-world processes and decision-making. Specifically, we investigate the surrogate modeling of three distinct examples. In the first two case studies, we trained the Bayesian LSTM models on data corresponding to a nominal set of physical properties such as thermal diffusivity and kinematic viscosity. Subsequently, we leverage these trained models to predict the system’s response when subjected to variations in this properties. Furthermore, in a third, distinct example, we extend our exploration to the realm of process systems, specifically Pressure Swing Adsorption (PSA) column. Here, we built a model trained on an extensive data set encompassing full operation cycles and by retraining the model using just one operation cycle with missing data, we successfully demonstrated its capability to extrapolate system behavior.