GitHub - Steel-predictor-project/Steel-predictor: Open-source ML model that predicts knife steel properties (toughness, edge retention, corrosion resistance, ease of sharpening) from alloy composition.

Steel Property Predictor — Composition-Based Modeling of Martensitic Alloy Steels

Predicts the mechanical and wear properties of high-carbon martensitic tool and cutlery steels directly from chemical composition, combining machine learning with physics-based metallurgical models trained on standardized laboratory test data.

134 alloys characterized across four property axes — wear resistance, impact toughness, corrosion resistance, and machinability — powered by CATRA wear measurements, Charpy impact data, and carbide-partition chemistry.

Interactive web tool — coming soon. This repository is the open-source model, data, and methodology.

How It Works

Most steel comparisons rely on one person's subjective opinions. This project is different — every property is derived from objective laboratory measurements and first-principles metallurgy:

The models predict base material properties on a 1–10 scale, then combine them with weighting profiles that reflect how different applications trade off those properties:

EDC:       40% edge retention + 30% corrosion + 25% toughness +  5% sharpening
Hard Use:  60% toughness + 25% edge retention + 10% corrosion +  5% sharpening
Kitchen:   45% edge retention + 35% corrosion + 10% sharpening + 10% toughness
Bushcraft: 55% toughness + 30% edge retention + 15% sharpening

Corrosion resistance is deliberately absent from Bushcraft: because the geometric mean floors each property at 0.5, even a 10% corrosion weight made a predicted 0.0 behave like a near-fatal flaw, which buried the carbon steels bushcraft knives are routinely made from. A bushcraft knife that needs oiling is normal.

The combination is a weighted geometric mean, so a steel that is unusable in one dimension cannot average its way to a good score. The result is rescaled onto 1–10 against the 111 knife steels in the dataset, where 10 is the best knife steel here for that use. The remaining 23 grades are hot-work, plastic-mould, holder and machinery steels that nobody builds knives from; they still get property predictions but are excluded from the use-case scale. See data/steel_applications.csv and docs/methodology.md.

Architecture

┌─────────────────────────────────────────────────────────────┐
│   RAW DATA                                                  │
│   Manufacturer PDFs (Crucible, Bohler, Carpenter, Hitachi)  │
│   Academic papers (CATRA tests, Charpy measurements)        │
│   → Compiled & normalized; every source cited in            │
│     DATA_SOURCES.md (raw extractions kept private)          │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│   FEATURE ENGINEERING                                       │
│   11 composition elements + 1 PM flag + 12 derived features │
│                                                             │
│   Derived: CVF, matrix Cr, PREN, Ms temp, VC fraction,     │
│   carbide former total, Cr/C ratio, PM interactions         │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│   MODELS                                                    │
│                                                             │
│   Edge Retention ─── XGBoost + RF + Ridge (0.45/0.30/0.25) │
│                      trained on CATRA TCC mm                │
│                      LOOCV MAE: 34.8mm (0.39 on 1-10)      │
│                                                             │
│   Toughness ──────── Ridge on physics features              │
│                      calibrated on Charpy ft-lbs            │
│                      Correlation: 0.96 (n=12)               │
│                                                             │
│   Corrosion ──────── Deterministic physics formula          │
│                      Matrix Cr + PREN + N contribution      │
│                      Validation r=0.89 vs KSN (n=61)        │
│                                                             │
│   Sharpening ─────── CVF + VC hardness (deterministic)      │
└───────────────────────────┬─────────────────────────────────┘
                            │
                            ▼
┌─────────────────────────────────────────────────────────────┐
│   OUTPUT                                                    │
│   Per-steel: 4 base scores + 4 use-case scores (1-10)      │
│   → data/processed/all_predictions.csv (134 steels)         │
│   → models/model_weights.json (serialized for inference)    │
└─────────────────────────────────────────────────────────────┘

Sample Output

Top steels by use case:

Full results for all 134 steels in data/processed/all_predictions.csv

Quick Start

# Clone the repo
git clone https://github.com/Steel-predictor-project/Steel-predictor.git
cd Steel-predictor

# Reproduce the model in one command (installs deps + runs the pipeline)
./run_training.sh

# ...or run the steps manually:
pip install -r requirements.txt
python scripts/train_model_v2.py   # trains from data/processed/training_ready.csv

Reproduces the edge-retention and toughness metrics exactly (edge-retention LOOCV MAE 0.391 on 48 CATRA steels; toughness r=0.96 on 12 Charpy steels) along with every prediction in all_predictions.csv.

Note on the corrosion r=0.89: the corrosion validation step compares our physics predictions against KnifeSteelNerds ratings. That reference set is third-party and is not redistributed with this repo, so the r=0.89 figure is not reproducible from a public clone — the validation is skipped automatically when the data is absent (you will see "KSN validation data not present"). The corrosion predictions themselves are a deterministic physics calculation (matrix Cr / PREN) and reproduce fully. See DATA_SOURCES.md and docs/methodology.md.

Output:

• data/processed/all_predictions.csv — scores for all 134 steels
• models/model_weights.json — serialized model for web inference
• models/model_summary.json — metrics and methodology summary

Requirements

• Python 3.10+
• scikit-learn ≥ 1.4
• XGBoost ≥ 2.0
• pandas ≥ 2.1
• numpy ≥ 1.26

Data Sources

Primary Measurement Data

Composition Data

Reference Ratings (for validation only)

• KnifeSteelNerds ratings used as a reference benchmark (not training target)
• Our model is methodologically independent — trained on machine measurements, not subjective ratings

Project Structure

steel-predictor/
├── scripts/
│   ├── train_model_v2.py      # Training pipeline (CATRA + Charpy + physics)
│   └── normalize_data.py      # Data preprocessing
├── data/
│   ├── processed/             # Normalized compilation + pipeline outputs
│   │   ├── all_predictions.csv
│   │   ├── training_ready.csv
│   │   └── unified_steels.json
│   └── LICENSE                # CC BY 4.0 (data + model)
├── models/
│   ├── model_weights.json     # Full serialized model (~27K lines)
│   ├── model_summary.json     # Metrics + methodology
│   └── xgb_catra_edge_retention.json  # Edge-retention ensemble (XGBoost)
├── scrapers/                  # PDF extraction scripts
├── docs/                      # GitHub Pages site
└── requirements.txt

Key Innovations

1. CATRA-trained edge retention — First open-source model using standardized machine cutting tests instead of subjective ratings
2. Carbide partition chemistry — Corrosion model accounts for V/Nb/W/Mo consuming carbon before Cr, preserving matrix chromium (critical for high-vanadium steels like S90V)
3. PM processing interactions — Models capture how powder metallurgy changes both toughness (finer carbides → less crack initiation) and corrosion (less Cr depletion at grain boundaries)
4. Application-weighted scoring — Output layer that weights the base properties differently per application profile rather than collapsing everything into a single "best steel" ranking

Contributing

Contributions welcome! Areas where help is needed:

• More CATRA data — Additional edge retention machine measurements improve the ML model
• Charpy data for conventional steels — Current toughness calibration is PM-only
• Heat treatment sensitivity — How do different HT parameters affect the same steel?
• International steels — Coverage of Chinese (e.g., SG2, VG-XTAL), Swedish (RWL-34), and German steels

How to contribute data

1. Fork the repo
2. Open an issue or PR adding your steel data with a public source citation (datasheet or published measurement). Raw source extractions are curated separately — see DATA_SOURCES.md for the sourcing standard.
3. New data is normalized into data/processed/ and the pipeline (scripts/train_model_v2.py) is re-run to validate.
4. All contributed sources are added to DATA_SOURCES.md.

License

This project uses two licenses:

• Code — Apache License 2.0 (see also NOTICE).
• Curated data & trained model — Creative Commons Attribution 4.0 (CC BY 4.0).

Attribution is required for reuse of the data or model: credit "Steel Property Predictor Project" with a link to this repository. These licenses cover only this project's own code, normalized compilation, derived features, and trained model. The underlying factual source data (compositions and lab measurements) is compiled from the third-party sources cited in DATA_SOURCES.md and remains the property of its respective publishers.

Links

• Interactive web tool: coming soon
• Methodology deep-dive: docs/methodology.md
• Project page: steel-predictor-project.github.io/Steel-predictor

Technical write-ups

• How We Built the Steel Property Predictor
• What Makes a Steel "Tough"? A Data-Driven Answer
• How Carbon Content Actually Affects Knife Performance
• Powder Metallurgy vs Conventional Steel: What the Numbers Say

An open-source research project. If this helped you understand or choose a steel, consider starring the repo.