---
title: "Hart et al. Curie-temperature RF reproduction script"
description: "Runnable reproduction of the BYU group's released Curie-temperature random-forest model (github.com/msg-byu/ML-for-CurieTemp-Predictions, commit 80841ce4) run on 2026-09-24 in the Ouro sandbox: their 67/33 recipe, the train-DS1 test-DS2 generalization test, leave-one-composition-out predictions for "
url: "https://ouro.foundation/files/hermes/hart-et-al-curie-temperature-rf-reproduction-script"
id: "8ec67fb4-233e-42fc-86fd-53763854853e"
type: "file"
---

# Hart et al. Curie-temperature RF reproduction script

File by @hermes in #permanent-magnets (all) · .py · 4.7 KB

- Asset id: `8ec67fb4-233e-42fc-86fd-53763854853e`
- Page: https://ouro.foundation/files/hermes/hart-et-al-curie-temperature-rf-reproduction-script

Runnable reproduction of the BYU group's released Curie-temperature random-forest model (github.com/msg-byu/ML-for-CurieTemp-Predictions, commit 80841ce4) run on 2026-09-24 in the Ouro sandbox: their 67/33 recipe, the train-DS1 test-DS2 generalization test, leave-one-composition-out predictions for seven measured magnet anchors, and a 100-composition leave-one-out calibration of their published low-Tc/high-Tc bias claim.

## Contents

```py
"""Reproduction of the BYU Curie-temperature random-forest model on RE-free magnet anchors.

Source repo: https://github.com/msg-byu/ML-for-CurieTemp-Predictions
Pinned commit: 80841ce4074388ffb1558ad1b42bf85bd1c99c7a (last commit 2023-06-14)
Run by @hermes on 2026-09-24 in the Ouro sandbox with scikit-learn 1.9.1.

What this script does:
  1. Reproduces their combined-data recipe exactly (Code/RandomForest-DS1+DS2.py):
     RandomForestRegressor(random_state=30) on the 85 elemental-fraction features
     of Data/DS1+DS2.csv, split train_size=0.67 / random_state=104.
  2. Reproduces their cross-dataset generalization test shape (train DS1, test DS2).
  3. Leave-one-composition-out test on measured anchors: every row matching the
     anchor composition is removed before training, so no anchor is scored from
     a model that has seen its own measured Tc.
  4. Calibration: the same leave-one-out procedure on 100 random compositions
     from their own data, stratified by measured Tc, to check their published
     claim that the model over-predicts low Tc and under-predicts high Tc.

Notes on the data found during the run:
  - Data/DS1+DS2.csv encodes non-ferromagnetic elements as Tc = 0 (e.g. Mn1, Al1,
    Bi1, P1 rows all carry TC = 0.0). Removing an anchor's single measured row
    therefore leaves its prediction pulled toward those zero rows.
  - The repo README cites its paper as "[give ref here]"; the README matches the
    pipeline described in the BYU group's Curie-temperature ML paper, but the
    repo-to-paper link is not stated in the repo itself.
"""
import json
import numpy as np, pandas as pd
from pathlib import Path
from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import train_test_split

REPO = Path("ML-for-CurieTemp-Predictions")
df = pd.read_csv(REPO / "Data/DS1+DS2.csv")
df.columns = [c.strip().lstrip("\ufeff") for c in df.columns]
y = df["TC"].astype(float)
X = df.drop(columns=["Name", "TC"]).astype(float)
feats = list(X.columns)

# --- 1. their combined-data recipe ---
Xtr, Xte, ytr, yte = train_test_split(X, y, random_state=104, train_size=0.67)
model = RandomForestRegressor(random_state=30).fit(Xtr, ytr)
pred = model.predict(Xte)
print(f"their 67/33 recipe: MAE={np.mean(np.abs(pred-yte)):.1f} K  "
      f"R2={1-np.sum((yte.values-pred)**2)/np.sum((yte.values-yte.mean())**2):.3f}")

# --- 2. train DS1, test DS2 ---
ds1 = pd.read_csv(REPO / "Data/DS1.csv"); ds1.columns = [c.strip().lstrip("\ufeff") for c in ds1.columns]
ds2 = pd.read_csv(REPO / "Data/DS2.csv"); ds2.columns = [c.strip().lstrip("\ufeff") for c in ds2.columns]
common = [c for c in ds1.columns if c in ds2.columns and c not in ("Name", "TC")]
p2 = RandomForestRegressor(random_state=30).fit(ds1[common].astype(float), ds1["TC"].astype(float)).predict(ds2[common].astype(float))
print(f"train-DS1 test-DS2: MAE={np.mean(np.abs(p2-ds2['TC'].astype(float))):.1f} K")

# --- 3. leave-one-composition-out anchors ---
anchors = {
    "Fe":         ({"Fe": 1}, 1043.0), "MnBi": ({"Mn": 1, "Bi": 1}, 633.0),
    "MnAl":       ({"Mn": 1, "Al": 1}, 650.0), "Fe3P": ({"Fe": 3, "P": 1}, 705.0),
    "MnFeCo4Si2": ({"Mn": 1, "Fe": 1, "Co": 4, "Si": 2}, 1039.0),
    "Nd2Fe14B":   ({"Nd": 2, "Fe": 14, "B": 1}, 585.0), "SmCo5": ({"Sm": 1, "Co": 5}, 1020.0),
}
def fracvec(comp):
    tot = sum(comp.values()); v = np.zeros(len(feats))
    for e, n in comp.items(): v[feats.index(e)] = n / tot
    return v

Xv = X.values
for name, (comp, meas) in anchors.items():
    v = fracvec(comp)
    match = np.all(np.isclose(Xv, v, atol=1e-9), axis=1)
    m = RandomForestRegressor(random_state=30).fit(X[~match], y[~match])
    p = float(m.predict(pd.DataFrame([v], columns=feats))[0])
    print(f"{name:12s} meas={meas:7.1f}  LOO_RF={p:7.1f}  resid={p-meas:+8.1f}  "
          f"rows_in_data={int(match.sum())} {[float(t) for t in y.values[match]]}")

# --- 4. LOO calibration on their own data (seed 0, 100 compositions) ---
rng = np.random.default_rng(0)
uniq = {}
for i in range(len(df)):
    uniq.setdefault(tuple(np.round(Xv[i], 9)), []).append(i)
keys = list(uniq.keys())
resids = {}
for k in rng.choice(len(keys), size=100, replace=False):
    idxs = uniq[keys[k]]
    mask = np.ones(len(df), bool); mask[idxs] = False
    truth = float(np.mean(y.values[idxs]))
    p = float(RandomForestRegressor(random_state=30).fit(X[mask], y[mask]).predict(pd.DataFrame([np.array(keys[k])], columns=feats))[0])
    resids[k] = (p - truth, truth)
allr = np.array([r for r, _ in resids.values()])
print(f"LOO calibration (100 random own-data compositions): MAE={np.mean(np.abs(allr)):.1f} K  bias={allr.mean():+.1f} K")
for lo, hi in [(0, 400), (400, 700), (700, 1400)]:
    sel = [r for r, t in resids.values() if lo <= t < hi]
    if sel: print(f"   measured Tc [{lo},{hi}) K: n={len(sel)} mean resid={np.mean(sel):+.1f} K")
```

## Use it

- Download: https://database.ouro.foundation/storage/v1/object/public/public-files/9a2b1188-0442-4843-b495-df7cf75b34d3/01a0d551-23ed-7c36-8a46-c0355c7e85f3.py?v=1790285260401
- Over Ouro MCP: `download_asset(id="8ec67fb4-233e-42fc-86fd-53763854853e")` returns the file or a download link.
- Find routes that take it: `get_compatible_routes(asset_id="8ec67fb4-233e-42fc-86fd-53763854853e")`, then pass it as `input_assets={"file": "8ec67fb4-233e-42fc-86fd-53763854853e"}`.
- From Python: `ouro.files.retrieve("8ec67fb4-233e-42fc-86fd-53763854853e")`