🔬 applied physics

NSF Future Manufacturing Data Challenge: A Multimodal DED Dataset for Probabilistic Local Geometry Prediction in Laser Tracks

This paper introduces a multimodal dataset for the NSF Future Manufacturing Data Challenge that integrates in-situ thermal imaging, scanning electron microscopy, and 3D optical profilometry of laser tracks on stainless steel 316L to enable probabilistic prediction of local geometric variations across four different laser power settings.

Original authors: Abhishek Hanchate, Himanshu Balhara, Satish T. S. Bukkapatnam

Published 2026-07-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Abhishek Hanchate, Himanshu Balhara, Satish T. S. Bukkapatnam

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a master chef use a super-hot laser wand to draw a single, perfect line of melted metal on a steel plate. You might think, "If I set the laser to the same power and move it at the same speed, the line will look exactly the same every time." But in the world of high-tech manufacturing, that's not quite true. The line wiggles, gets wider, or gets bumpy in tiny, unpredictable ways.

This paper introduces a brand-new "recipe book" (a dataset) for scientists and computer programmers to help them predict exactly how that laser line will look, down to the very smallest bumps. It's like giving a detective three different types of clues to solve a mystery about why a laser track turns out the way it does.

The Three Clues (The Data)
To solve the mystery of the laser line, the researchers gathered three different kinds of "eyes" to watch the process:

The Heat Camera: They used a special camera that sees heat, snapping 50 pictures every second. These pictures show the glowing "melt pool"—the tiny, liquid puddle of metal right in front of the laser. The camera sees a square area about 5.6 mm by 5.6 mm. Even though the puddle is small, its shape, brightness, and how it cools down tell a story about what's happening.
The Microscope (SEM): After the laser finished its job, they took super-close-up pictures of the steel plate using a scanning electron microscope. Think of this as looking at the ground after a car has driven over it to see what the terrain was like underneath. Is the ground smooth? Are there tiny hills or valleys? These pictures show the "terrain" the laser drove over, providing context for the local surface features.
The 3D Ruler (Height Map): After the laser finished its job, they used a fancy 3D scanner to measure the final line. This creates a map showing exactly how high or low every tiny spot is, measured in micrometers (millionths of a meter).

The Big Challenge
The goal of this challenge is to teach a computer to look at the heat pictures and the "ground" pictures and then guess what the final 3D shape will look like.

Here is the tricky part: The paper explicitly says do not just guess the average width of the whole line. The laser line isn't a perfect, straight ruler; it's a living, breathing thing that changes as it moves. The computer needs to predict how the width or shape changes at every single spot along the line, like predicting a bumpy road rather than just saying "the road is flat."

What They Did (and Didn't Do)
The team ran experiments on stainless-steel plates using four different laser power settings: 200, 300, 350, and 400 Watts. They moved the laser at a steady speed of 10 mm per second. They focused on a specific 80 mm stretch of the line (from the 20 mm mark to the 100 mm mark) because that's where all three cameras could agree on the location.

They are very careful about one rule: The computer is not allowed to peek at the final 3D map to make its guess. That would be like cheating on a test by looking at the answer key. The computer must only use the heat and ground clues to predict the future shape.

How Sure Are They?
The authors have collected real data from actual machines, not just computer simulations. They have provided the raw files, some starter code, and a guidebook for anyone who wants to try solving this puzzle. However, they aren't claiming to have solved the problem yet. Instead, they are saying, "Here is the data; we suggest that if you combine these three clues, you might get a much better prediction than before."

They even set aside one specific track (Track 21) as a special test case. This track is a bit messy because the scanner had a glitch and didn't capture the whole height map perfectly. They want to see if the computer can still make a good guess even when the final picture is incomplete.

The Takeaway
This isn't a magic wand that instantly fixes manufacturing. It's a toolkit. The paper suggests that by combining how the metal heats up, what the surface looks like underneath, and how the final shape turns out, we can start to understand the tiny, local reasons why laser lines aren't perfect. It's a step toward making future manufacturing smarter, but the heavy lifting of building the perfect prediction model is up to the people who use this new dataset.

Technical Summary: NSF Future Manufacturing Data Challenge Dataset for Probabilistic Local Geometry Prediction

Problem Statement
Laser-based Directed Energy Deposition (DED) produces tracks with spatially varying geometries. While global process parameters influence the final shape, the local geometry is also driven by the dynamic thermal history of the melt pool and the pre-existing state of the substrate (e.g., local valleys, ridges, and ripples). Current modeling approaches often rely on single global width values, failing to capture these local dependencies. The challenge is to explicitly model the relationship between process-driven thermal dynamics, substrate morphology, and the resulting local geometric variation (such as width, boundary position, or contour deviation) along the scan direction.

Methodology and Dataset Construction
The paper introduces a multimodal dataset designed to support the NSF Future Manufacturing Data Challenge. The dataset was generated using an Optomec LENS® MTS 500 platform performing single-track "bead-on-plate" scans on SS316L substrates without powder feed, isolating laser-material interaction from powder-flow effects.

The experimental design involves four laser power settings (200, 300, 350, and 400 W) at a fixed scan speed of 10 mm/s. The dataset integrates three complementary modalities, all aligned to a common physical coordinate window of 20–100 mm:

In-situ Thermal Imaging: Captured via a Stratonics ThermaViz sensor at 50 fps. Each 400×400 pixel frame (14 µm/pixel resolution) covers a 5.6 mm × 5.6 mm field of view. The data captures the melt pool's size, shape, intensity, asymmetry, and cooling tail, which drive process-dependent variation.
Scanning Electron Microscopy (SEM): Acquired using a Zeiss EVO MA10 system. These images provide context regarding the surrounding substrate morphology (valleys, ridges, ripples). To prevent data leakage, the processed track region is masked or excluded when using SEM features for prediction. The dataset includes 13–14 tiles per track, with Tile 01 corresponding to the 100 mm coordinate and higher-numbered tiles corresponding to the 20 mm coordinate.
Post-process Height Maps: Generated using a Bruker ContourGT-K white-light 3D optical profilometer. Stored in Wyko ASCII format, these maps provide full-field height data (x, y in mm; z in nm) from which local geometric descriptors are extracted.

Challenge Task
Participants are tasked with modeling a probabilistic local geometry descriptor $p(g_i(x) | T_{i,t-k:t+k}, S_i, x)$ , where $x$ is the physical scan-direction coordinate. The inputs are a sequence of thermal frames ( $T$ ), an SEM context image ( $S$ ), and the target is a local geometric descriptor ( $g$ ), such as local width, boundary position, or edge roughness. The primary expected target is local width variation extracted from the height map.

Key Contributions

Multimodal Alignment: The paper establishes a coordinate convention that links thermal sequences, SEM tiles, and height maps over a specific physical interval (20–100 mm), addressing the difficulty of aligning disparate sensor data in DED processes.
Dataset Release: The release includes four specific track conditions (IDs 8, 10, 14, and 21) with synchronized data. It provides starter notebooks, participant-facing code, and reusable loader functions to facilitate data ingestion and coordinate transformation.
Explicit Local Dependencies: By providing high-resolution substrate context (SEM) alongside thermal history, the dataset encourages models to move beyond global averages and predict spatially varying geometric quantities.

Results and Evaluation
As this is a dataset release paper for a challenge, it does not present predictive model results or benchmark performance. Instead, it outlines the evaluation framework:

Validation Strategy: Due to the limited number of tracks (four), the paper recommends cross-track validation, specifically treating Track 21 as a held-out test case. Track 21 is noted to have less complete post-process profilometry due to acquisition failure, making it suitable for testing model robustness to incomplete data rather than for hyperparameter tuning.
Metrics: Suggested metrics include Mean Absolute Error (MAE) for local width, boundary-position error, and probabilistic metrics such as Negative Log Likelihood (NLL) or Continuous Ranked Probability Score (CRPS) for uncertainty-aware outputs.

Significance
The paper positions this dataset as a foundational resource for the NSF Future Manufacturing Data Challenge. Its significance lies in enabling the development of models that combine process-driven (thermal) and substrate-driven (morphological) sources of variation. By making local dependencies explicit and providing a multimodal, physically aligned dataset, the authors aim to advance the state of the art in predicting local geometric quality in DED processes, moving from global parameter estimation to spatially resolved probabilistic modeling.

Technical Summary: NSF Future Manufacturing Data Challenge Dataset for Probabilistic Local Geometry Prediction

More like this