Intelligent prediction of Si3N4 ceramic strength considering inherent defect characteristics

The complex, randomly distributed pores in ceramics cause significant variability in fracture strength, limiting reliable applications. This study proposes an intelligent method for quantitative strength prediction of Si3N4 ceramics, achieving a validation error of ∼2 %. The excellent predictive performance of the model was partly attributed to the precise extraction of pore size, shape, orientation, and location obtained through nano-CT technology. Additionally, the construction of a scientific feature space (based on correlation analysis and recursive elimination methods) enabled the optimal construction of strength prediction models based on XGBoost and AdaBoost algorithms from 9 classic machine learning algorithms. More importantly, the model comprehensively integrated the inherent properties of the material and pore feature information, realized through hyperparameter tuning, region-based pore screening, and consideration of critical flaw characteristics of the material. Interpretable machine learning (SHAP, PDP and ICE), non-destructive testing, and fracture analysis results were also used for descriptor extraction and model analysis.