A data-driven framework for prediction and inverse design of propionic acid reactive extraction

Propionic acid can be recovered from aqueous streams using reactive extraction, but finding the best operating conditions usually requires a lot of laboratory work. With the least amount of experimentation, this study creates a data-driven framework that predicts extraction efficiency and suggests ideal conditions. In order to find combinations that reach a target efficiency, machine-learning models were trained on a carefully selected dataset of propionic acid systems and combined with inverse design. Eight regression models: Artificial Neural Networks, Random Forest Regressor, XGBoost, CatBoost, Gradient Boosting, k-Nearest Neighbors Regressor, Huber Regressor, and Linear Regression were trained. Benzene, toluene, paraffin liquid, petroleum ether, methyl isobutyl ketone (MIBK), 1-octanol, 2-octanol, hexanol, hexane, sunflower oil, kerosene, heptane, oleyl alcohol, ethyl acetate, butyl acetate, and 1-dodecanol are among the chemically varied classes represented in the solvent set. Aliquat 336, tri-octylamine (TOA), and tributyl phosphate (TBP) are extractants used in the studies. Explainable Artificial Intelligence uncovered solvent-specific contributions that affect efficiency and explained why extractant concentration dominates performance. Thousands of design options were quickly narrowed down to a shortlist of candidates that achieve the target. Several inverse-design suggestions were experimentally verified and delivered extraction efficiencies close to the intended 90% efficiencies, thereby offering experimental support to the predictive and inverse-design framework. This approach offers an achievable method to lower experimental load while enhancing decision quality, providing a transferable strategy to speed up advancements in reactive extraction and associated separation techniques.