Machine learning modeling of the volumetric overall mass transfer coefficient in a novel L-shaped pulsed sieve plate extraction column

The volumetric overall mass transfer coefficient is a key design parameter in liquid–liquid extraction columns, yet reliable prediction remains difficult for non-standard geometries. This study applied machine learning to simultaneously predict the volumetric overall mass transfer coefficient for both the horizontal and vertical sections of a novel L-shaped pulsed sieve plate extraction column — a task not addressed by any single correlation, since the two sections operate under distinct dispersion and mixer-settler hydrodynamic regimes — using artificial neural network, random forest, and support vector regression models trained on experimental data from two liquid–liquid systems. The artificial neural network model achieved the highest accuracy for both sections, with a test coefficient of determination above 0.98 and an average absolute relative error below 6% on a small hold-out set (four test conditions per section), outperforming published empirical correlations developed for conventional vertical columns. Sensitivity analysis identified dispersed phase holdup and drop size as the dominant predictive variables in both sections. Neural-network-based response surface analysis identified optimal operating conditions, with residual diagnostics showing no evident systematic bias across the range of predicted values. Given the limited dataset (36 operating conditions per section), the model is regarded as valid for interpolation within the investigated experimental domain, and its generalization capability remains to be confirmed with larger datasets.