Explainable Artificial Intelligence in modelling hydrogen gas solubility in n-Alkanes

• Explained AI model predicts hydrogen solubility in n-alkanes accurately. • XAI techniques optimize inputs and show data anomalies. • Linear relationship found between solubility and key variables. • Anomalies in C1 require specialized modelling considerations. This research aims to enhance the predictive modelling of hydrogen gas (H 2 ) solubility in n-alkanes using Explainable Artificial Intelligence (XAI) techniques. The focus is on elucidating the impact of key variables on solubility, optimizing model inputs, and ensuring data integrity. The study employed the Extra Trees (ET) regression model complemented by XAI approaches, including Partial Dependence Plots (PDP), Individual Conditional Expectation (ICE) plots, and Friedman H-statistics for assessing feature interactions. Feature importance (FI) was quantified using Permutation Feature Importance (PFI), Tree-based Feature Importance (TFI), and Partial Dependence Feature Importance (PDFI), which facilitated informed feature selection and model refinement. Analysis revealed that pressure ( P ), dimensionless P ( PD ), and dimensionless temperature ( TD ) significantly influence H 2 solubility, demonstrating a near-linear relationship. The application of XAI not only optimized model inputs but also played a critical role in identifying and correcting data anomalies, enhancing overall data quality. The refined model, ET2_4, demonstrated improved accuracy, achieving a Root Mean Squared Error ( RMSE ) of 0.0085 on testing data, with H-statistics confirming strong interactions, particularly between P , PD , and TD . Notable deviations were observed for C1, suggesting specialized modelling considerations for atypical n-alkanes. The integration of XAI techniques provided profound insights into variable interactions and solubility dynamics, significantly advancing the accuracy of predictive models for H 2 solubility ( x ) in n-alkanes. The findings emphasize the necessity of incorporating advanced analytical methods in chemical process simulations to ensure data reliability and model efficacy. Future research should explore alternative characterization methods to extend these insights across diverse chemical systems, especially for compounds exhibiting deviated behaviours.