Prediction of residual stress distribution induced by ultrasonic nanocrystalline surface modification using machine learning

Ultrasonic Nanocrystalline Surface Modification (UNSM) offers an efficient and cost-effective approach for enhancing material mechanical properties by inducing Severe Plastic Deformation (SPD). It leads to grain refinement and substantial residual stress generation beneath the workpiece surface. This study investigates the influence of key modification parameters, specifically static load, vibration amplitude, and strike tip size on compressive residual stress (CRS) distribution. A Finite Element Method (FEM)-based model for the UNSM process is developed, and validated against experimental outcomes, yielding a dataset of 45 unique cases across various modification scenarios. The Balancing Composite Motion Optimization (BCMO), as a meta-heuristic algorithm is used to optimize the hyperparameters of the Support Vector Regression (SVR) model. Additionally, the performance of Artificial Neural Network (ANN), Polynomial Chaotic Extension (PCE), and Kriging algorithms is evaluated in parallel. Among these Machine Learning (ML) models, the SVR-BCMO emerges as a pioneer for its accuracy in estimating residual stress. A sensitivity analysis employing Sobol’ indices further clarifies the distinct impact of each input parameter on residual stress distribution resulting from UNSM. In essence, this research offers a tool for rapidly estimating residual stress, even in cases of limited datasets. Furthermore, the findings help in making prompt decisions regarding of UNSM conditions. This is achieved by elucidating the effect of each input parameter and facilitating the determination of residual stresses in specific scenarios.