AI-based surrogate modelling for incremental sheet metal forming using large-scale dataset

• A fully automated workflow is proposed for large-scale finite element data generation in incremental sheet metal forming. • A trajectory-aware AI surrogate based on a Transformer Cross-Attention U-Net is introduced for forming predictions. • The model encodes full tool-path histories and integrates them with spatial deformation features through cross-attention. • Both final and time-resolved sheet geometries are predicted directly from forming trajectories and die shapes. • Results demonstrate high predictive accuracy and an order-of-magnitude speedup compared to high-fidelity FEM simulations. Incremental Sheet Metal Forming (ISMF) enables the manufacture of complex, customised metal components without dedicated tooling, yet accurate finite element simulation of the process remains computationally intensive. To overcome this limitation, we present a fully automated digital workflow that integrates large-scale physics-based simulation with a deep learning surrogate model tailored specifically to ISMF. Thousands of high-fidelity simulations were generated using an automated pipeline for synthetic die shape construction, Z-constant tool path generation, numerical input assembly, and structured post-processing. A controlled explicit dynamic formulation with stabilised time integration was employed to reduce simulation runtime while preserving the essential deformation behaviour. To enable rapid prediction of forming outcomes, we introduce the Transformer Cross-Attention U-Net (TCAU), a neural architecture that represents the evolving tool path as a sequence of latent trajectory tokens and fuses them with spatial U-Net features through multi-head cross-attention. The model is trained in a non-autoregressive manner using distributed GPU computation, allowing efficient learning from large-scale spatio-temporal datasets. TCAU accurately predicts both intermediate and final sheet geometries and provides speed-ups of several orders of magnitude relative to full finite element simulations. The proposed framework enables fast ISMF evaluation, supports accelerated tool path optimisation, and offers a scalable foundation for data-driven die shape compensation.