Deep-learning-aided modifier adaptation: synergies with process intensification

Abstract

Deep learning allows for functions, and their gradients, to be approximated to a high accuracy. Modifier adaptation is a real-time optimization method, which is used to optimize process economics online, and requires gradients to make first-order model corrections. In this work, backpropagated gradients are computed from neural networks trained on historical steady-state data, thus not explicitly requiring any gradient data for training. Data curation and convergence properties are discussed for the proposed method. The deep-learning-aided modifier adaptation is tested in analogous simulated integrated and intensified reactor-separator systems, where it is shown to reconcile plant and model optima in the presence of model mismatch. The case studies show better economics and constraint satisfaction when using the intensified system and the deep-learning-aided modifier adaptation. Further, intensification and deep-learning-aided modifier adaptation are observed to work in tandem as both accelerate the convergence of the plant to its true optima. The proposed method shows how historical data logs can be leveraged to address epistemic uncertainty and improve performance in model-based optimization, especially in intensified systems.