Inteligência artificial para previsão de variáveis de processos termomecânicos via elementos finitos

Abstract

Metal extrusion is widely used in the manufacturing of components with complex geometries and high mechanical performance requirements. However, predicting critical process variables - such as maximum load and stress triaxiality - through Finite Element Method (FEM) simulations, like those performed using QForm UK, involves high computational costs and limited generalization to new alloys. To address these challenges, this study proposes the use of machine learning algorithms as an alternative to predict such variables from the chemical composition of metallic alloys. Simulations were carried out using aluminum (Al), copper (Cu), nickel (Ni), and stainless steel (Fe) alloys under fixed process conditions. The extracted outputs were maximum load and maximum triaxiality, while the input variables were physicochemical descriptors derived from the chemical compositions, such as atomic mass, atomic radius, and electronegativity. These features were used to train three supervised regression models: Support Vector Regression (SVR), Random Forest (RF), and Gradient Boosting Regressor (GBR). The models’ hyperparameters were automatically optimized using the Optuna library, which applies a Bayesian search strategy more efficient than traditional methods like grid search. The GBR model achieved the lowest Root Mean Square Error (RMSE) in predicting load, particularly for copper alloys, although it struggled more with nickel and aluminum. For triaxiality, all models showed limited performance, with errors comparable to the variable’s natural variability. The results confirm the feasibility of integrating FEM simulations with artificial intelligence for load prediction in extrusion processes, while also highlighting the challenges in modeling stress triaxiality.