Hardness prediction in the upsetting process of Al%ZrO2--an approach to machine learning using regression and classification models

The current study focuses on the prediction of metal hardness distribution in upsetting tests for different compositions of ZrO2 embedded with an aluminum matrix using machine learning algorithms and finite element (FE) analysis. The mass fraction of the ZrO2 particles varied from 4% to 8%, and three sets of solid cylindrical rods with Al4%ZrO2 , Al6%ZrO2 , and Al8%ZrO2 were prepared using the stir casting method. The upsetting process was simulated, and an equation for predicting hardness was developed from the equivalent strain distributions. Artificial neural networks (ANNs), multilinear regression (MLR) along with equations developed from FE analysis were used to train the model for regression analysis, considering the principal stresses, friction factor, anisotropy ratio, effective strain, and hoop strain as input and the magnitude of hardness as output parameters. Regression analysis reveals that ANN (tri-layer network), XGBoost, and MLR algorithms are the best suitable for the given data sets with a root mean square (R2) greater than 0.95. XGBoost, ANN (narrow), and SVM are linear and are the most recommendable classifier algorithms for the current investigation. Hardness data from ring compression tests were used to validate the results obtained from the trained models with the test results.