Machine Learning Based Approach for Prediction of Hood Oilcanning Performances

Computer Aided Engineering (CAE) simulations are an integral part of the product development process in an automotive industry. The conventional approach involving pre-processing, solving and post-processing is highly time-consuming. Emerging digital technologies such as Machine Learning (ML) can be implemented in early stage of product development cycle to predict key performances without need of traditional CAE. Oil Canning loadcase simulates the displacement and buckling behavior of vehicle outer styling panels. A ML model trained using historical oil canning simulation results can be used to predict the maximum displacement and classify buckling locations. This enables product development team in faster decision making and reduces overall turnaround time. Oil canning FE model features such as stiffness, distance from constraints, etc., are extracted for training database of the ML model. Initially, 32 model features were extracted from the FE model. Domain expertise and variable selection techniques were implemented to clean up the database for dependencies and duplicates. This resulted in identification of 21 key parameters for training the ML model. Database for buckling classification model is highly skewed with only 5% data points with buckling. Synthetic data is generated using SMOTE algorithm to overcome data imbalance. These features are then used to train and validate the ML model for buckling. Predictive model developed using Extreme Gradient boosting (XG Boost) algorithm with R² more than 90% for training and test datasets. It predicted maximum displacement with 20% error for 80% test data points. Also, buckling data points are classified with 98% accuracy. Prediction made using the ML model is in good agreement (< 20% error) with CAE results. This resulted in substantial time savings from 11 days to 30 minutes for the prediction of key performances.