Prediction of forming limit diagrams using machine learning

Measuring forming limit diagrams (FLDs) is a time consuming and expensive process. Data-driven and artificial intelligence approaches have been used to facilitate finding relationships between alloy composition, and manufacturing processing conditions with mechanical properties of materials. Machine learning (ML) methods are a promising route to predict FLD of aluminium alloys that are increasingly used in automotive structural applications. In the present work, we developed a machine learning (ML) based tool to establish the relationships between alloy composition / thermomechanical processing route to the material’s FLD, which is a measure of formability. Measured FLDs of various 5XXX and 6XXX aluminium alloys along with their chemical composition and thermomechanical processing parameters like homogenization, hot and cold rolling parameters etc. and the respective mechanical properties like n and r values were manually curated and stored in a database. A two-stage ML model was trained on this database. In the first stage, the minimum and maximum points of the minor strain were predicted using a feature set incorporating chemical composition, thermomechanical processing parameters, and n and r values using support vector regression (SVR). In the second stage, the predicted minor strain in the first stage was used as input in addition to the same feature set to predict the major strain using gradient boost regression (GBR). The trained ML model successfully predicted FLDs with R 2 value above 0.93. To study the impact of the difference between a predicted and a measured FLD on stamping simulation results, finite element simulations of a cross die were performed. The difference in draw depth of the cross-die using the predicted FLD and the measured FLD was between 5 to 10%. The current results show that ML can be a viable way for predicting FLDs.