Machining processes induce residual stresses near the surface of a workpiece, significantly affecting its accuracy, performance, and service life. MISUTECH’s MISULAB software predicts these stresses via two coupled FEM simulations: ABAQUS (2D, tool-material thermomechanical loads) and Sysweld (3D, residual stress projection). Each cycle takes approximately 8 hours, with the Sysweld phase being the main computational bottleneck, limiting fast exploration of tool geometries. Hence, this thesis aimed to develop an ML-based reduced-order model to predict residual stress profiles. The scope of the study encompassed longitudinal turning of 15-5PH martensitic stainless steel. The input features of the model comprised three cutting conditions (Vc, f, ap) and three parameters of tool geometry (αn, γn, Rβ). The targets were derived by parametric curve fitting of the simulated residual stress profiles using the Bernstein Polynomial and Gaussian-Exponential models. According to the results, Gaussian Process Regression was the best-performing regressor for the Bernstein Polynomial, while performance under the Gaussian Exponential parametrisation was best achieved by Support Vector Regression in the cutting direction and Gaussian Process Regression in the feed direction. Overall, the proposed model has strong potential as a rapid surrogate tool for predicting residual stresses, reducing computational time and resource requirements while maintaining high predictive performance.
|
