1 May 2026 in SSRN Electronic Journal
Background: Selecting optimal extractant solvents for liquid-liquid extraction (LLE) of polycyclic aromatic hydrocarbons (PAHs) requires balancing interdependent physicochemical factors across 20 priority analytes and multiple solvent classes. Systematic experimental screening is time-and resource-intensive, while published machine learning approaches have not combined quantum-chemical descriptors with stacking ensemble architectures for this task. Results: A two-level stacking ensemble was developed, integrating gradient boosting (CatBoost), ensemble decision trees (Random Forest (RF)), and a hybrid multilayer perceptron with bidirectional long short-term memory (MLP+BiLSTM) as base learners, with a linear meta-model trained via 5-fold out-of-fold (OOF) cross-validation. The feature space comprised 18 physicochemical and DFT-derived quantum-chemical descriptors. Across 10 independent runs, CatBoost delivered the highest predictive accuracy (R² = 0.896 ± 0.032, mean absolute error (MAE) = 8.77 ± 1.58%) and lowest variability, outperforming MLP+BiLSTM by ΔR² = 0.109. The stacking ensemble exceeded the best direct-application baseline by ΔR² = +0.117 on a challenging data split (p = 2.4∙10^(-5)). Solvent physicochemical descriptors-solvation cavity volume (15.8%), dielectric constant (14.4%), and boiling point (13.3%)-collectively dominated feature importance (57.4%), consistent with solvation theory. Significance: The proposed methodology enables in silico screening of candidate solvents, reducing experimental effort by an estimated 70-80% during analytical method development. The stacking framework with OOF cross-validation is transferable to other solvent-selection tasks in analytical chemistry, including mobile-phase optimisation in chromatography and sorbent design for solid-phase extraction.