HUGO-CS: A Hybrid-Labeled, Uncertainty-Aware, General-Purpose, Observational Dataset for Cold Spray

This paper introduces HUGO-CS, a large-scale, hybrid-labeled observational dataset of 4,383 cold spray experiments derived from scientific literature using a novel uncertainty-aware framework that combines automated LLM extraction with targeted manual refinement to overcome data scarcity and standardization challenges in process optimization.

The Gist

Imagine the world of Cold Spraying as a high-stakes cooking competition. In this kitchen, chefs (scientists) use a special technique to build metal objects layer by layer without melting them, kind of like using a super-fast, high-pressure air cannon to shoot tiny metal particles at a surface so they smash together and stick.

The problem is that every chef has their own recipe. Some write their recipes in a secret code, some use different units of measurement (cups vs. grams), and many just scribble the results in a messy notebook with no clear list of ingredients. Because of this, it's incredibly hard to figure out the "perfect recipe" for building strong, durable metal parts.

Here is what this paper does, explained simply:

1. The Problem: A Library of Messy Notebooks

For years, scientists have been publishing papers about cold spraying. But if you wanted to learn from all of them, you'd hit a wall:

• The Data is Hidden: The results are often trapped inside pictures or tables in PDF files, not in a format a computer can easily read.
• The Scale is Tiny: Previous attempts to collect this data were like trying to build a house with only a few bricks. The biggest collection before this had only 137 experiments.
• The Inconsistency: One paper might say "Aluminum 6061," another "AA 6061," and a third "Al 6061 Powder." To a computer, these look like three completely different materials, even though they are the same.

2. The Solution: The "HUGO" Chef's Assistant

The authors built a new system called HUGO (Hybrid-labeled, Uncertainty-aware, General-purpose, Observational) to fix this. Think of HUGO as a super-smart, tireless robot assistant who helps a team of human chefs organize the library.

• The Robot (LLM): They used a Large Language Model (a type of AI) to read thousands of scientific papers and pull out the numbers. The robot is fast—it can read a paper in seconds.
• The Safety Net (Human Review): Robots make mistakes. Sometimes they hallucinate (make things up) or miss details hidden in a chart. So, the authors didn't just trust the robot. They created a "Risk Mitigation" system.
  • Imagine the robot is sorting mail. If the envelope looks weird, the robot puts it in a "Red Bin."
  • Humans then only open the "Red Bin" to fix the mistakes.
  • If the envelope looks normal, the robot keeps it.
  • This saves time because humans only check the tricky stuff, not every single paper.

3. The Result: The "HUGO-CS" Cookbook

The result of this process is a massive new dataset called HUGO-CS.

• Size: It contains 4,383 experiments from 1,124 different papers. That is 30 times bigger than any previous collection.
• Detail: It tracks 144 different features for every experiment, from the type of gas used to the exact shape of the metal powder.
• Cleanliness: The team cleaned up the data. They turned "Al 6061," "AA 6061," and "Aluminum 6061" all into one standard label. They also converted different units (like inches vs. millimeters) so everything speaks the same language.
• The Gold Standard: Out of the 4,383 experiments, 1,765 were double-checked by humans. This creates a "Gold Subset" that researchers can trust completely to test their own theories.

4. What They Did With It

The paper shows that this new, clean cookbook actually works. They used it to train computer models to predict how strong a metal part will be.

• They successfully predicted the strength of aluminum alloys.
• They successfully predicted the hardness of various metal powders.
• Crucially, they found that knowing the exact chemical recipe (composition) of the powder was the most important factor for making accurate predictions.

5. The Takeaway

This paper didn't invent a new way to spray metal. Instead, they built the ultimate library for people who study metal spraying. By combining a fast robot with smart human checks, they turned a chaotic pile of messy scientific notes into a clean, organized, and massive dataset that anyone can use to understand and improve cold spray technology.

In short: They took a messy, fragmented library of 1,000+ books, used a robot to read them, had humans fix the robot's mistakes, and turned it all into one giant, perfectly organized encyclopedia for metal builders.

Technical Summary

Technical Summary: HUGO-CS and the HUGO Framework

Problem Statement
Cold spraying is a versatile solid-state additive manufacturing process with significant applications in component repair and manufacturing. However, optimizing this process is hindered by the complexity of interdependent parameters and a critical lack of large-scale, machine-readable data. While scientific literature contains numerous relevant experiments, results are inconsistently reported (often embedded in tables and figures), utilize non-uniform units, and are frequently tailored to specific material classes or narrow target properties. Existing curated datasets are small (the largest prior collection contained only 137 experiments) and often lack the granularity required for high-performing predictive modeling or broad process optimization. Furthermore, manual extraction of data from literature is prohibitively time-consuming, averaging 91 minutes per document, which limits the scale of available data.

Methodology: The HUGO Framework
To address these limitations, the authors introduce HUGO (Hybrid-labeled, Uncertainty-aware, General-purpose, Observational), a framework designed to extract structured experimental records from scientific literature at scale. The methodology combines the speed of Large Language Models (LLMs) with targeted manual verification through a Hierarchical Risk Mitigation (HRM) strategy.

1. Text Extraction and Preprocessing: Structured PDFs are converted to machine-readable Markdown using MinerU to preserve layout and table structures. Metadata is retrieved via the Crossref API, with manual intervention for ambiguous cases.
2. Schema Construction: A comprehensive schema of 144 features was developed, covering material properties, experimental parameters (e.g., carrier gas, powder morphology), and testing conditions.
3. Hybrid Extraction Pipeline:
   • LLM Labeling: An instruction-based LLM (GPT o4-mini) performs zero-shot inference to extract experimental data into a structured JSON format.
   • Hierarchical Risk Mitigation (HRM): Instead of uniform random sampling, HRM flags high-risk outputs for manual review in four stages:
     • Structural Errors: Identifying unparsable JSON or truncated responses.
     • Completeness Errors: Detecting schema non-compliance (missing or extra fields) and attempting automated correction via string-similarity matching before manual re-labeling.
     • Statistical Outliers: Flagging content-based anomalies using domain-informed thresholds, global outliers (deviation >3σ from dataset mean), and local outliers (deviation >2σ from material class centroid).
     • Coverage Errors: Estimating the number of experiments reported in figures versus text. A secondary LLM prompt estimates the expected number of experiments; articles with a high gap between expected and extracted counts (weighted by metric rarity) are prioritized for manual labeling.
4. Post-Processing and Standardization:
   • Categorical Consolidation: A "Propose-Inspect-Review" (PIR) workflow merges free-text aliases (e.g., "Al 6061" vs. "AA 6061") into standardized categorical values.
   • Continuous Composition Mapping: Feedstock chemistries are mapped to a structured 50-element representation, normalizing units (e.g., converting at.% to wt.%) and handling powder blends.
   • Unit Normalization: Numeric values are standardized (e.g., MPa, GPa), and hardness values are parsed to separate the unit from the test load.
   • Provenance Labeling: A secondary labeling pass identifies results derived from non-standard testing methods (e.g., nanoindentation vs. ASTM E8 tensile testing) to flag potential comparability issues.

Key Contributions

1. HUGO Framework: A novel hybrid-labeling workflow that balances LLM efficiency with manual accuracy by prioritizing manual intervention for high-risk extractions.
2. HUGO-CS Dataset: A large-scale, machine-readable dataset containing 4,383 cold-spray experiments with 144 features extracted from 1,124 primary sources. This represents a 30x increase over the previous largest dataset (137 samples).
3. Gold Subset: A high-fidelity, hand-labeled subset of 1,765 experiments from 243 sources, intended for benchmarking, error analysis, and training high-fidelity models.
4. Standardization Pipeline: An extensive cleaning pipeline that consolidates categorical descriptors, maps chemistries to continuous compositions, and normalizes units across diverse sources.
5. Performance Evaluation: A critical assessment of LLM extraction performance, identifying common failure modes such as inability to parse figures, grouping errors, and verbosity issues.
6. Open Access: Full release of the dataset and pipeline code under a CC-BY license to support reproducibility and extension to new domains.

Results

• Scale and Coverage: The final dataset spans diverse materials (Al, Cu, Ti, etc.) and processing conditions, capturing a wide range of mechanical properties including microhardness (2,980 values), yield strength (506 values), and porosity.
• Extraction Accuracy: On a held-out validation set of 20 articles (80 ground-truth experiments), the LLM achieved a precision of 89.61% and a recall of 86.25%. Feature-level accuracy averaged 94.55%.
• Error Analysis: Primary failure modes included the inability to extract data reported exclusively in figures and errors in grouping experimental conditions (e.g., splitting tensile and hardness results into separate records). The HRM strategy successfully identified and corrected structural, schema, and coverage errors, significantly improving data quality.
• Modeling Proof-of-Concept: Two models were trained to demonstrate utility:
  • An aluminum yield strength model (58 samples) achieved an $R^2$ of 0.66 and MAE of 36.6 MPa using Gradient Boosting.
  • A multi-material microhardness model (2,431 samples) achieved an $R^2$ of 0.65 and MAE of 87.12 HV using CatBoost. Feature importance analysis highlighted the value of detailed feedstock composition data.

Significance
The paper claims that HUGO-CS significantly expands the scale and accessibility of experimental cold-spray data, enabling modern data-driven analysis and predictive modeling that were previously constrained by small, fragmented datasets. By providing a general-purpose, uncertainty-aware, and standardized dataset, the work facilitates meta-analysis, process optimization, and the development of robust predictive models. The authors emphasize that the inclusion of detailed feedstock compositions and the standardization of units and categories allow for richer analysis, such as studying alloying effects and transfer learning, which were difficult with prior reductive datasets. The release of the HUGO framework also offers a pathway for extending this approach to other scientific domains requiring large-scale literature extraction.