Predicting the mechanical and wear properties of martensitic alloy steels from chemical composition
Most steel comparisons come down to one person's subjective opinions. This project is built on objective laboratory measurements — CATRA standardized wear tests, Charpy impact data, and first-principles carbide chemistry.
XGBoost + RandomForest + Ridge ensemble trained on 48 CATRA TCC machine measurements. LOOCV MAE: 34.8mm (0.39 on 1-10 scale).
Physics-informed Ridge regression calibrated against 12 Charpy impact tests. Captures carbide type effects (VC vs Cr7C3) and PM processing.
First-principles calculation of effective matrix chromium after carbide partition. Accounts for V/Nb scavenging carbon before Cr.
Carbide hardness model based on CVF and VC/WC content relative to sharpening stone hardness (alumina 2100 HV).
| Steel | Tough | Edge | Corr | Sharp | Best For |
|---|---|---|---|---|---|
| CPM MagnaCut | 6.6 | 4.2 | 6.3 | 1.9 | All-rounder |
| CPM 3V | 8.0 | 3.4 | 2.9 | 5.5 | Hard Use |
| Vanax | 5.1 | 4.1 | 10.0 | 8.0 | Kitchen |
| CPM S35VN | 4.5 | 4.1 | 6.9 | 2.9 | EDC |
| M390 | 3.1 | 5.3 | 8.9 | 1.8 | Slicing EDC |
Full results for 134 steels available in the predictions CSV.
48 steels tested with standardized machine cutting protocol by Larrin Thomas (2020). Objective edge retention measurement.
Chemical compositions extracted from Crucible, Bohler-Uddeholm, Carpenter, Hitachi/Proterial, and Alleima technical data sheets.
12 PM steels with room-temperature Charpy V-notch testing. Used to calibrate the physics-based toughness model.
KnifeSteelNerds ratings used as independent validation benchmark (not training target). Our model is methodologically independent.
Longer write-ups on the model and the metallurgy behind the predictions:
How the XGBoost + RF + Ridge ensemble works, feature engineering, and small-data ML with 48 samples.
Why carbide type — not carbide volume — decides whether a knife edge chips.
More carbon does not mean better edge retention. What carbon actually does, from the data.
Same composition, made two ways: exactly what PM processing buys you.
Read the full methodology documentation for details on each model.