NSF releases multimodal DED dataset linking thermal imaging to 316L track geometry at four laser powers
A new public multimodal DED dataset pairs thermal video, SEM, and height maps for 316L laser tracks at four powers. It enables probabilistic local geometry prediction models directly applicable to production cost reduction. Evidence strength is limited by single-substrate, single-speed experiments; broader parameter sweeps are required for generalization.
The dataset captures single-bead DED scans at fixed 10 mm/s speed using Stratonics ThermaViz thermal video, Zeiss EVO SEM imagery, and Bruker white-light profilometry over aligned 20-100 mm windows. Starter code and coordinate conventions enable direct fusion of melt-pool dynamics with final bead height and width distributions. This supports probabilistic models that predict local geometric variation rather than mean dimensions. Industrial DED adoption for repair and near-net-shape parts hinges on reducing post-machining allowances; geometry prediction errors above 50 µm currently drive scrap rates above 15 % in production cells. Public release of synchronized thermal-to-height data removes a key barrier for machine-learning teams targeting that tolerance.
NSF Challenge organizers: At least three teams will report <40 µm median absolute error on held-out tracks within nine months of dataset release.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2607.07965)
- [2]Supporting Source(https://doi.org/10.1016/j.addma.2023.103456)