SHAP enhanced machine learning prediction of structure–property relationships in plasma sprayed lanthanum zirconate coatings for TBC applications
来源：X-MOL
The performance limitations of traditional yttria-stabilized zirconia (YSZ) have prompted the investigation of lanthanum zirconate (LZ) thermal barrier coatings (TBCs) as potential alternative topcoat materials. This study addresses the limitations of conventional statistical methods by introducing an innovative hybrid framework that integrates the Design of Experiments (DoE), Machine Learning (ML), and SHapley Additive exPlanations (SHAP) to model and understand the process-structure-property relationships in plasma-sprayed LZ coatings. A Central Composite Design technique was employed to conduct 20 experimental trials by systematically varying the plasma power, spray distance, and powder flow rate. Analysis of variance (ANOVA) revealed that the plasma power had a significant effect on the coating characteristics (p < 0.0001). Advanced ML models demonstrated superior predictive accuracy, and the XGBoost algorithm accurately predicted coating microhardness (R2 = 0.989). SHAP analysis provided quantitative, model-independent interpretability, revealing that plasma power (mean |SHAP| value = 121.8) had over 40 % more impact on microhardness than other parameters. A strong quantitative relationship was established between porosity and microhardness, showing a ∼33 HV increase in microhardness for every 1.0 vol % reduction in porosity. The framework was thoroughly validated, and the prediction errors were limited to ±5 %. This integrated DOE-ML-SHAP approach offers an innovative, interpretable, and data-driven model for advancing and optimizing next-generation TBCs.