Graph convolutional network for axial concentration profiles prediction in simulated moving bed

The simulated moving bed (SMB) chromatographic separation is a continuous compound separation process based on the differences in adsorption capacity exhibited by distinct constituents of a mixture on the fluid phase and stationary phase. The prediction of axial concentration profiles along the beds in a unit is crucial for the operating optimization of SMB. Though the correlation shared by operating variables of SMB has an enormous impact on the operational state of the device, these correlations have been long overlooked, especially by the data-driven models. This study proposes an operating variable-based graph convolutional network (OV-GCN) to enclose the underrepresented correlations and precisely predict axial concentration profiles prediction in SMB. The OV-GCN estimates operating variables with the Spearman correlation coefficient and incorporates them in the adjacency matrix of a graph convolutional network for information propagation and feature extraction. Compared with Random Forest, K-Nearest Neighbors, Support Vector Regression, and Backpropagation Neural Network, the values of the three performance evaluation metrics, namely MAE, RMSE, and R², indicate that OV-GCN has better prediction accuracy in predicting five essential aromatic compounds' axial concentration profiles of an SMB for separating p-xylene (PX). In addition, the OV-GCN method demonstrates a remarkable ability to provide high-precision and fast predictions in three industrial case studies. With the goal of simultaneously maximizing PX purity and yield, we employ the non-dominated sorting genetic algorithm-II optimization method to perform multi-objective optimization of the PX purity and yield. The outcome suggests a promising approach to extracting and representing correlations among operating variables in data-driven process modeling.