Interpretable machine learning-based prediction of hydrodynamic performance of divided fixed-valve tray in distillation columns

Abstract

Accurate prediction of tray hydrodynamics is essential for the design and optimization of modern distillation equipment. Conventional empirical correlations often show limited generalizability across diverse operating conditions, restricting their practical applicability. In this study, a data-driven modeling (DDM) framework integrated with Interpretable machine learning (ML) was developed to predict the hydrodynamic characteristics of a divided fixed-valve tray, including wet pressure drop, clear liquid height, weeping rate, and entrainment rate. A total of 760 experimental data points were obtained from cold-model experiments. Four machine learning algorithms—Random Forest, Extreme Gradient Boosting, Support Vector Regression, and a shallow Multilayer Perceptron—were trained and compared. Model performance was evaluated using the coefficient of determination (R²), root mean square error (RMSE), and a composite R²-based index (R_com). Among these, the SVR model achieved the highest predictive accuracy, with R_com values of 0.9750 and 0.9834 for the test and extension sets, respectively. Building upon the best-performing SVR model, an expanded feature set was constructed using physically informed transformations of the orifice kinetic energy factor and overflow strength. Pearson correlation and SHAP analyses were applied to quantify each feature’s contribution, revealing the dominant hydrodynamic drivers and clarifying their positive or negative influences on tray behavior. The integrated modeling and interpretability framework not only enhances predictive reliability but also provides actionable insight for guiding tray design and operational optimization.