A modular transfer learning approach for complex chemical process network modeling

Abstract

Traditional first-principles modeling often struggles to capture complex interactions in chemical process networks, while machine learning methods require large datasets, which are often unavailable. To address these issues, we propose a modular transfer learning framework that integrates modules developed for individual processes into a global model for the entire process network. Three types of modules are used: black-box neural networks for processes with sufficient data, a foundation module based on the reptile method for data-limited cases, and a first-principles module for processes where the physicochemical phenomena are well understood. These modules are integrated using an aggregation module to capture the nonlinear dynamics and complex interactions within the process network. The effectiveness of the proposed approach is demonstrated on a polymer production process modeled in Aspen Plus Dynamics, where it significantly outperforms conventional fully-connected neural network models in accuracy, flexibility, and computational efficiency.