Discover the SciOpen Platform and Achieve Your Research Goals with Ease.
Search articles, authors, keywords, DOI, etc.
The interaction model parameters are the key thermodynamic parameters in activity coefficient models such as Wilson and non-random two-liquid (NRTL) models. Their accuracy directly determines the reliability of the results of vapor – liquid phase equilibrium and separation process calculations. The core bottlenecks of traditional parameter acquisition methods relying on experimental data regression and universal functional activity coefficient (UNIFAC) model estimation include limited data coverage, unstable solution of nonlinear equations and limited scope of application. To address these issues, we constructed a dual-branch graph neural network (GNN) model based on molecular structure characterization enabling efficient and direct prediction of the parameters of the Wilson interaction model. This model takes molecular diagrams as input, separately extracts the molecular features of solutes and solvents, and, through the message-passing mechanism, deeply learns the potential structural correlations between molecules. At the same time, we introduce a parallel residual strategy to reduce the prediction error. Using 4068 binary system interaction model parameter data for training and testing, our proposed parallel residual GNN model significantly outperforms the UNIFAC method across multiple error indicators, and compared with the single GNN model, the prediction accuracy is further improved. The predicted parameters Aij, Aji, Bij, Bji of the Wilson interaction model were substituted into the Wilson equation to calculate the infinite dilution activity coefficient γi∞. The results were highly consistent with the calculated value in the literature, verifying the thermodynamic rationality and engineering applicability of our model.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Comments on this article