Integrating van't Hoff Equation with Artificial Neural Network for the Prediction of H₂S Solubility in Ionic Liquids

Accelerating the discovery of high-performance ionic liquids (ILs) for hydrogen sulfide (H₂S) capture is hindered by the prohibitive cost of experimental screening and the limited generalization capability of purely data-driven machine learning models, which often lack thermodynamic consistency and fail in complex chemical extrapolation tasks. To address these challenges, this study introduces a physics-informed hybrid artificial neural network (hybrid ANN) that integrates the van't Hoff equation directly into the model architecture. This approach preserves the data-driven flexibility of machine learning while enforcing fundamental thermodynamic constraints, ensuring that predictions remain physically consistent across unexplored chemical spaces. The hybrid ANN demonstrated exceptional predictive accuracy, achieving a coefficient of determination (R²) exceeding 0.9987 on a conventional test set. More critically, in a rigorous generalization test involving unseen cation-anion combinations, the purely data-driven ANN model suffered a catastrophic performance collapse, with R² plummeting from 0.9990 to 0.8215, whereas the hybrid ANN maintained robust performance with an R² of 0.9925. This highlights the essential role of embedded physical knowledge in safeguarding model reliability under data-sparse conditions. Interpretability analysis using SHAP revealed that structural features, particularly the imidazolium cation and alkyl chain substituents, exert a stronger influence on H₂S solubility than operational pressure, underscoring the dominance of molecular design over process parameters. The model further captured smooth, continuous nonlinear interactions among cations, anions, and side chains, providing actionable insights for the rational design of next-generation IL-based H₂S capture solvents. This work establishes a paradigm for integrating domain knowledge with deep learning to enable reliable, interpretable, and generalizable predictive modeling in complex chemical systems.