NSF Future Manufacturing Data Challenge: A Multimodal DED Dataset for Probabilistic Representation and Prediction of Laser-Track Geometry
arXiv:2607.07965
The paper presents a multimodal dataset for directed energy deposition (DED) that combines in‑situ thermal images, SEM images, and 3D height maps to enable probabilistic prediction of local geometry variations in laser‑track builds on stainless‑steel 316L.
Abstract
We introduce a multimodal directed energy deposition (DED) dataset for predicting the probabilistic local geometric variation of single laser tracks produced on stainless-steel 316L substrates. The dataset supports the NSF Future Manufacturing Data Challenge and contains three complementary modalities: in-situ thermal image sequences from a Stratonics ThermaViz melt-pool sensor, scanning electron microscopy (SEM) images acquired using a Zeiss EVO MA10 system, and full-field height maps acquired using a Bruker ContourGT-K white-light 3D optical profilometer. Each experiment is a bead-on-plate scan at one of four laser powers, 200, 300, 350, and 400 W, with a fixed scan speed of 10 mm/s. The release includes a multimodal coordinate convention linking thermal, SEM, and height-map measurements over a common physical 20--100 mm window, with the raw dataset available on Zenodo and participant-facing notebooks, reusable code, and documentation available on GitHub.
4 pages, 2 figures