Reliable modeling of sulfur solubility in sour gas mixtures using advanced ensemble learning frameworks

Accurate prediction of sulfur solubility in sour gas mixtures is critical for mitigating operational risks associated with sulfur deposition in pipelines and processing facilities. In this study, we present a new machine learning (ML) modeling framework combined with post-hoc interpretability analysis to model sulfur solubility across a wide range of thermodynamic and compositional conditions using a comprehensive experimental database. Three robust ML models, including Power Law Ensemble Method (PLEM), Extra Trees (ET), and Generalized Regression Neural Network (GRNN), were trained and evaluated to capture complex nonlinear relationships governing sulfur solubility in sour gas mixtures. Among the developed paradigms, the PLEM approach achieved superior predictive performance, with R 2 values of 0.9999 (training) and 0.9821 (testing), and RMSEs of 0.0001 and 0.0428, respectively. Over 94% of PLEM predictions had an absolute error below 0.05, demonstrating high reliability. To enhance model transparency and support physical validation, post-hoc interpretability analysis was conducted using SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-Agnostic Explanations). These analyses were used to interrogate trained model behavior and assess the contribution of input variables, confirming physically consistent trends. Outlier detection using the leverage method further validated data reliability and safeguarded against anomalous predictions. By combining high predictive accuracy with post-hoc interpretability of model behavior, this framework provides a reliable data-driven tool for real-time monitoring and decision support in sour gas processing, while improving confidence in model deployment for safety-relatwd operations. Beyond operational efficiency, the proposed methodology contributes to risk reduction and safer design, aligning with current priorities in process safety under the era of AI.