MACHINE LEARNING ENABLED INVESTIGATIONS FOR FORMABILITY PREDICTION OF SHEET METALS

Abstract

The objective of this thesis is to determine formability of sheet metal using onset of necking and fracture during forming process. Material properties from research papers were collected and feature engineered to remove missing values, duplicate entries, and outliers. Following this the labelled datasets has been utilised to train various supervised ML models viz SVR, RFR, KNN, DTR and XGB. Performance metrics and coefficients of determination were subsequently used to assess the performance of various models. Following the development of ML models, the second phase involves conducting experimental stretch forming on EDD 1 mm thick automobile grade sheet metal. An in-house stretch forming setup was used to deform samples in three distinct strain paths: biaxial, plane strain, and uniaxial. The blank underwent deformation until the appearance of necking. The FLD predictions of the available ML models were compared to the experimental results, and only Random Forest regressor accurately predicted onset of necking in EDD. Furthermore, ML models have been utilised in numerical methods to determine the limiting dome height of a stretched sheet metal. The Yld2000-2d yield model was used to determine dome height, which was then compared to experimental values. And found that all ML models in the uniaxial strain path can predict within 15% of the mean absolute percentage error, but the biaxial and plane strain paths have very high error rates. Following the completion of the forming limit diagram, the next stage of research was to find the fracture forming limit diagram (FFLD), using same method used for FLD machine learning model. The hyperparameters of ML models have been tuned to improve their accuracy. Experimental material properties derived from a researcher's work were utilised to forecast the FFLD for the purpose of assessing the performance of the ML model. RFR had R2 0.9173 and MAPE 12.082%, while SVR had R2 0.6285 and MAPE 20.069% making it the best and worst sheet metal onset of necking prediction model. The best and worst models to predict sheet metal forming fracture were RFR with R2 0.9512 and MAPE 12.15% and KNN with R2 0.8196 and MAPE 25.2%.