Research on Predicting Gas–Solid Erosion in Series Elbows Using Interpretable Machine Learning

Abstract

Erosion in industrial pipelines is inevitable, making accurate prediction essential for ensuring equipment safety. This study employs interpretable machine learning models to predict erosion in serial elbows under gas–solid flow conditions. A predictive model was developed by integrating computational fluid dynamics (CFD) with the Euler–Lagrange method. Latin hypercube sampling (LHS) was applied to five key factors influencing pipeline erosion rates (ER). Five tree-based ensemble machine learning models were selected, optimized using grid search, and subsequently employed to predict the wall-averaged and maximum erosion rates at both upstream and downstream elbows in serial pipelines. To analyze feature interactions, correlation analysis, Shapley Additive Explanations (SHAP), and response surface methods were utilized. Results indicate that the optimized CatBoost model demonstrated high accuracy in predicting gas–solid erosion in serial elbows, while SHAP analysis enhanced model interpretability. In combination with correlation and response surface analyses, both qualitative and quantitative evaluations of factor interactions were conducted. This study improves the predictive capability and interpretability of industrial pipeline erosion modeling, offering valuable insights for erosion prevention and control in industrial applications.