Challenge Instructions

Challenge Prompt

Given a sequence of thermal images recorded during laser scanning, develop a model that predicts the local geometric variation of the final laser track.

Participants should treat the laser track as a spatially varying feature rather than a perfectly uniform line. The objective is to learn how patterns in the thermal images—especially around the moving melt pool—relate to the final track geometry measured after processing. The prediction may focus on local track width, boundary position, contour irregularity, or other descriptors that capture how the track changes along its length.

Main Goals:

1. Extract thermal descriptors of the moving melt pool

Develop image-based features that describe the melt pool and its surrounding thermal field. These descriptors may include melt pool size, shape, temperature distribution, intensity gradients, asymmetry, cooling-tail behavior, frame-to-frame changes, or learned features extracted using machine-learning models.

2. Represent laser track variation as a spatial signal

Instead of describing each track using only one average width value, represent the final track geometry as a quantity that changes along the scan direction. This representation may include local width, boundary position, boundary fluctuation, edge roughness, waviness, or local contour deviation.

3. Predict local track geometry from thermal history

Use one thermal frame, or a short sequence of consecutive frames, to predict the corresponding local segment of the final laser track. The model may make predictions at the frame level, over short spatial windows, or over longer track segments.

4. Account for multiple sources of variation

The final track variation may come from both the laser scanning process and the original condition of the substrate. Some variations may be caused by changes in heat input, melt pool shape, or local process behavior, while others may be influenced by surface or subsurface variations already present in the plate. Participants are encouraged to develop models that can identify, or quantify these different sources of variation where possible.

5. Provide interpretable links between thermal behavior and final geometry

Prediction accuracy is important, but the model should also provide insight. Participants should explain which thermal features are most strongly related to track formation, local geometric variation, or irregular track boundaries.

Challenge Instructions

PDF of Intro PowerPoint

About the Data

The dataset contains thermal image sequences collected during laser scanning at four different laser powers (200 W, 300W, 350W, 400W). In each experiment, the laser creates a 100 mm long track on a stainless-steel 316L plate. The laser moves at a scan speed of 10 mm/s, and the thermal camera records images at 50 frames per second. Therefore, each consecutive thermal frame corresponds to approximately 0.2 mm of laser travel.

Each thermal frame is 400 × 400 pixels, with a physical resolution of approximately 14 µm per pixel. The melt pool is located near the center of each frame. The temperature field of the melt pool changes as the laser moves, providing a time-resolved view of how heat is distributed during track formation.

After laser scanning, the final tracks were measured using profilometry. From these height maps, geometric information can be extracted along the scan direction, including local track width, left and right boundary positions, edge irregularity, and local contour variation. These profilometer-derived measurements should serve as the ground truth for model training and evaluation.