A machine learning-based surrogate model for fast residual stress prediction in cold spray additive manufacturing

This study presents a data-driven surrogate strategy for efficient residual stress prediction in the cold-spray additive manufacturing. A comprehensive dataset was generated from thermomechanical finite element simulations, with residual stress values extracted as targets. Using key process parameters—nozzle travel speed, heat flux, and substrate thickness—several algorithms, including Random Forest, Extra Trees, and XGBoost, were evaluated. The Extra Trees model achieved the best performance, with the highest coefficient of determination ( R ²) and the lowest MAE, MAPE, and RMSE. Model interpretability was examined via SHapley Additive exPlanations (SHAP), indicating that element position played a leading role. The surrogate model was further validated by predicting residual stress distributions in components with increased deposition layers. Overall, the framework provides a fast and reliable alternative to computationally intensive simulations, enabling real-time process optimization and design exploration in cold-spray additive manufacturing.