Machine learning-based residual stress prediction in shot-peened components: Application to additively manufactured surfaces with stochastic roughness

Shot peening is an efficient technique to improve the fatigue behaviour of metal components. In this study a hybrid approach, combining numerical and experimental is presented to investigate the influence of initial surface roughness on the distribution of residual stress in shot-peened components with particular attention to additively manufactured (AM) parts characterized by significant surface roughness (Ra=10−35µm). A 3D Finite Element (FE) model incorporating real rough surface morphology and strain-rate sensitive material model with nonlinear kinematic/isotropic hardening was developed, and subsequently was validated with experimental data available. A comprehensive methodology was employed, beginning with the development of a finite element model to simulate single and multi-random shot impacts on smooth and artificially generated AM-representative rough surfaces, while including stochastic surface topographies commonly seen in AM processes. Results indicate that SP decreased the roughness on surfaces with a high initial average roughness, Ra = 35.4 μm, by 16.67 %. However, the average surface roughness on samples that had lower roughness originally experienced an increase by 18 %. The results showed that the high roughness decreases the maximum and depth of the compressive residual stresses induced by SP, in comparison with smooth sample. Additionally, Machine Learning (ML) algorithms, Artificial Neural Networks and Random Forest, were applied to predict RS distribution based on SP parameters and initial surface roughness topography. The ML models were trained using a large dataset of all applicable process variables, which gives a significant predictive framework for SP applications.