Quantification of the influence of laser shock peening parameters on residual stresses in thick AA 7050 specimens: An experimental and explainable machine learning-based approach

Laser Shock Peening (LSP) is a mechanical treatment capable of inducing a compressive residual stress field in the surface of a component, thereby improving its fatigue behavior. However, the selection of process parameters and the component geometric factors, such as thickness, are of paramount importance in the optimization of the final residual stresses, which exhibit a complex response due to the interaction between the process variables. An explainable machine learning framework has the potential to model the LSP treatment with high accuracy, while concurrently providing interpretations on the results and the process parameters' influence and interrelation. In this study, the effect of LSP on residual stresses in thick AA 7050-T7451 was experimentally investigated using incremental hole drilling with a Gaussian Process Regression-based approach to minimize user influence. Subsequently, experimental data was used to construct a robust ensemble machine learning model, namely a Random Forest, which achieved high generalization power by utilizing a rigorous K-Folds cross-validation with 100 repetitions. Finally, a critical comparative analysis of different approaches to determine feature importance was conducted. In this context, the SHapley Additive exPlanation (SHAP) framework emerged as a significant tool, providing reliable results and interpretations at both global and local levels. In this way, a quantitative assessment of the contributions of the LSP variables for each individual residual stress measurement was achieved. This allowed a deeper understanding of the effects of these parameters and their mutual interactions. • Hole Drilling residual stress measurements in thick Laser shock peened aluminum samples • Gaussian Process Regression is used to quantify uncertainty in residual stresses • Random Forest modeled Laser shock peening achieving high generalization power • Explainable AI using SHAP values quantified parameters’ influence and interactions