Comparative evaluation of neural network models for geometric deviation prediction and toolpath compensation in double-sided incremental forming

Double-sided incremental forming (DSIF) is a dieless sheet metal forming process suitable for low-volume production but is limited by springback-induced geometric deviation. In this work, machine learning is used to predict the geometric deviation of a formed part. Then, these values are applied to the initial computer-aided design (CAD) model as toolpath compensation to provide a cost-effective approach to improve the accuracy of the final part compared to the target geometry after unclamping and springback. Three neural network models, i.e., a feedforward artificial neural network (ANN), a convolutional neural network (CNN), and a graph neural network (GNN), are trained on a unified experimental dataset with deep drawing quality (DDQ) steel as the material formed and directly compared for geometric deviation prediction and accuracy after toolpath compensation. The dataset is generated by forming 25 parts with diverse geometries and measuring their post-unclamping geometric deviations. Although model-specific preprocessing is required, all neural network models use consistent local geometric input features, e.g., spatial coordinates and surface normal components, and the corresponding measured geometric deviation as the output. Two unseen geometries are used for evaluation. The ANN, CNN, and GNN models achieve average prediction mean absolute errors of 0.45, 0.32, and 0.41 mm, respectively, on the two unseen geometries. The predicted deviations are then used to create modified CAD models, reducing the geometric deviation by 34%, 49%, and 51% for ANN-, CNN-, and GNN-based compensation, respectively. This study shows that neural network models with explicit spatial information aggregation, i.e., CNN and GNN models, outperform the ANN model in improving the geometric accuracy of DSIF.