Integrated design, control, and machine learning‐based composition prediction for extractive distillation using deep eutectic solvent

Abstract Accurate, real‐time composition prediction in complex distillation systems is essential for advanced process control and optimization. This study presents an integrated framework combining process design, dynamic control, and machine learning‐based modelling for an extractive distillation column. A choline chloride‐glycerol deep eutectic solvent (DES) used as green entrainer for separating a ternary azeotropic mixture of ethyl acetate, ethanol, and water. Two configurations are investigated: a conventional triple‐column extractive distillation (TCED) system and a modified four column extractive distillation (FCED) design incorporating a pre‐concentrator. Both systems are optimized using a sequential iterative methodology, with FCED demonstrating enhanced performance achieving a 32.3% reduction in total annual cost (TAC) and a 6.78% decrease in CO 2 emissions. To ensure operational robustness, two control strategies (CS1 and CS2) are implemented in the FCED system using Aspen Dynamics. CS2, employing a temperature–composition cascade control structure, significantly improves ethanol purity control under feed disturbances. For real‐time composition prediction, random forest (RF)‐based machine learning models are trained on 2000 closed‐loop data samples. 17 critical process variables are selected as inputs to predict the compositions of ethyl acetate, ethanol, and water. All RF models achieve R 2 values above 0.99 and exhibit extremely low error metrics on both training and test sets. Learning curve analysis confirms their generalization capability, validating their effectiveness for real‐time monitoring and intelligent control of complex distillation systems.