Surrogate Modeling of Catalytic Packed Bed Reactor Using Deep Learning Architectures and Its Application in Modeling Catalyst Performance

Abstract The field of machine learning is evolving rapidly, and deep learning–based surrogate modeling has emerged as an efficient approach for accelerating numerical studies. In this work, multiple deep learning architectures were investigated as surrogate models for simulating methane oxidation in a packed-bed reactor. Among the tested architectures, a Long Short-Term Memory (LSTM)-based model demonstrated superior performance on unseen data and was able to capture the underlying relationships within the data set. The surrogate model was subsequently employed for a gradient-based optimization task to accurately simulate the deactivation behavior of palladium-based catalysts when exposed to oxidizing, humid conditions.