ML Prediction of Magnet Curie Temperature

ISEF Category: Materials Science

Subcategory: Electronic, Optical, and Magnetic Materials  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Permanent magnets power motors, speakers, and wind turbines. The catch is that many strong magnets rely on rare earths, which are expensive and supply-chain fragile. You can study which compositions stay magnetic at high heat, without making magnets in a lab. That makes this a great way to do real materials research from a laptop.

What Is It?

Curie temperature is the point where a magnetic material stops acting like a permanent magnet. Think of it like the moment a crowd loses its rhythm. Below that temperature, the spins inside the material stay aligned well enough to keep magnetism. Above it, thermal motion scrambles that order.

Your project asks whether a machine learning model can predict Curie temperature from compositional descriptors. Descriptors are numbers that describe a material’s ingredients, such as atomic radius, electronegativity, or elemental fractions. Instead of testing one magnet at a time in a lab, you train a model on known data and ask it to spot patterns in chemistry. Tools like matminer help turn formulas into features a model can read.

Why This Is a Good Topic

This makes a strong science fair topic because you can test a clear prediction problem with public data and real metrics like error and R². You also connect to a real engineering problem, finding magnets that keep working at high temperatures without rare earth elements. You can learn data cleaning, feature engineering, model comparison, and validation, which are useful skills in materials science and computer science.

Research Questions

• How does the choice of compositional descriptors affect Curie temperature prediction accuracy?
• What is the effect of removing rare-earth-containing compounds from the training set on model performance?
• Does a tree-based model outperform linear regression for predicting Curie temperature from composition?
• To what extent does adding elemental statistics improve predictions compared with using stoichiometry alone?
• Which descriptor groups contribute most to prediction error for rare-earth-free magnets?
• How does the model perform on compounds with compositions unlike the training data?

Basic Materials

• Laptop with internet access.
• Spreadsheet software for tracking data cleaning steps.
• Python installed on a computer.
• Jupyter Notebook or Google Colab.
• matminer package for descriptor generation.
• pandas for data handling.
• scikit-learn for model building and evaluation.
• Public dataset of magnetic materials with Curie temperature values.

Advanced Materials

• Access to a larger curated materials database, such as Materials Project-derived datasets.
• Python environment with matminer, pandas, scikit-learn, numpy, and matplotlib.
• SHAP or another feature-interpretation library.
• Hyperparameter tuning tools such as GridSearchCV or Optuna.
• Version-controlled code repository for tracking experiments.
• Optional access to a workstation for larger feature sets and repeated cross-validation.

Software & Tools

• Python: Runs your data cleaning, feature engineering, and model training workflow.
• Jupyter Notebook: Lets you document code, plots, and notes in one place.
• matminer: Converts chemical formulas into compositional descriptors.
• scikit-learn: Trains and evaluates regression models with cross-validation.
• pandas: Organizes the dataset and makes cleaning easier.
• matplotlib: Plots prediction error, feature importance, and model comparisons.

Experiment Steps

1. Define the target variable and decide exactly which compounds belong in your dataset.
2. Clean the public data so you remove duplicates, missing values, and inconsistent formula records.
3. Choose a first set of compositional descriptors and compare them with a simpler baseline feature set.
4. Train several regression models and use the same validation strategy for each one.
5. Test whether the model still works when you hold out rare-earth-free compounds or unusual compositions.
6. Interpret the strongest features and connect them back to chemistry instead of treating the model like a black box.

Common Pitfalls

• Mixing data from different sources without checking whether the Curie temperature values were measured the same way.
• Letting near-duplicate compounds appear in both training and test sets, which inflates accuracy.
• Using too many descriptors for too little data, which makes the model memorize noise.
• Ignoring class imbalance or composition imbalance, which can make the model look better than it is on rare-earth-free magnets.
• Reporting only one metric, which hides whether the model makes large errors on the compounds that matter most.

What Makes This Competitive

A competitive version of this project goes beyond building one model. You compare descriptor sets, test multiple validation schemes, and explain why the model succeeds or fails on rare-earth-free compounds. Strong entries also use error analysis, not just a single score, to show where the method breaks. If you can connect the model’s top features back to magnet chemistry, your work looks much more like real materials research.

Project Variations

• Focus only on rare-earth-free ferrites and see whether the model improves when you narrow the chemical space.
• Compare compositional descriptors against crystal-structure descriptors if you can find matched data.
• Predict another thermal property, such as magnetic anisotropy proxy values, and compare how hard each target is to model.

Learn More

• matminer documentation: Read the package guide and examples on the matminer project documentation site.
• Materials Project: Search the public materials database for composition and property data.
• Open Quantum Materials Database: Explore another public source for materials-property data.
• scikit-learn User Guide: Find regression, cross-validation, and model evaluation guides on the scikit-learn documentation site.
• PubMed: Search for review articles on machine learning in magnetic materials and Curie temperature prediction.
• npj Computational Materials: Read open papers on data-driven materials discovery through the journal website and library access.