Image-based Artificial Intelligence-driven modelling for blank shape optimisation in sheet metal forming

• Non-parametric shape optimisation using image-based Artificial Intelligence models. • High-fidelity image-based AI surrogate model and differentiable shape generator. • Systematic accuracy comparison between image-based and scalar-based AI surrogate models. • Demonstration on an industrial problem containing multi-parameterised geometries. • Automated optimisation that shortens design iterations from weeks to minutes. Design for manufacturing is essential to fully exploit the potential of emerging materials and processing technologies. However, traditional trial-and-error optimisation often exhibits inferior performance in manufacturability-driven problems, particularly when handling complex shapes. Surrogate modelling and optimisation have been widely investigated for efficiently predicting simulation results and enhancing manufacturability. Nevertheless, existing methods are mostly constrained by fixed shape parameterisation schemes, limiting their flexibility and effectiveness. To overcome this limitation, this research develops a non-parametric optimisation framework, validated on a sheet metal forming case study, specifically the blank shape optimisation of a hot-stamped B-pillar. The framework integrates an auto-decoder, serving as a differentiable blank shape generator, a convolutional neural network (CNN)-powered surrogate model for manufacturability evaluation, and an Adam optimiser for automated shape optimisation. The surrogate model predicts thickness distributions from the signed distance fields (SDFs) of blank shapes, which are generated by the auto-decoder from latent vectors; based on the predictions, the optimiser iteratively updates the latent vectors to acquire a blank shape with optimised manufacturability. The proposed framework demonstrates superior performance in terms of the accuracy of thickness distribution prediction, the fidelity of blank shape generation, and the efficiency of blank shape optimisation.