Open challenges for learning-based manufacturability in sheet metal bending

Develop and evaluate learning methods for intra-process manufacturability assessment in sheet metal bending that (i) capture long-range geometric dependencies inherent to bending manufacturability, (ii) generalize across different punch and die tooling configurations, and (iii) transfer effectively from synthetic benchmarks to real-world deployment.

Background

The paper introduces a taxonomy for manufacturability metrics and the BenDFM synthetic dataset for sheet metal bending, then benchmarks state-of-the-art geometric deep learning models. Results show topology-aware models outperform point-cloud models, yet configuration-dependent targets remain challenging.

In concluding remarks, the authors explicitly identify key open challenges that must be addressed to advance deployment: capturing long-range geometric dependencies involved in bending, generalizing across different tooling configurations, and bridging the gap between synthetic datasets and real-world manufacturing scenarios.