Bayesian Design of Thermoelectric Oxides

ISEF Category: Materials Science

Subcategory: Computation and Theory  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Some materials can turn wasted heat into electricity. That sounds simple, but finding the right recipe is hard because tiny changes in composition can change performance a lot. Bayesian optimization helps you search smarter, not harder. You can use it to predict which high-entropy oxides might make better thermoelectrics.

What Is It?

This project studies how to use active learning and Bayesian optimization to design new oxide materials. Think of it like a smart guessing game. Instead of testing every possible composition, you let the computer pick the next best candidate based on what it already learned from earlier results.

A thermoelectric material turns a temperature difference into electric voltage. For a good thermoelectric, you want low thermal conductivity, which means heat does not move through it very easily. High-entropy oxides are a family of materials made from several different elements mixed into one crystal structure. The mix can scatter heat-carrying vibrations, called phonons, which may lower thermal conductivity.

Bayesian optimization is a search method that balances exploration and prediction. It uses a model to estimate which materials look promising, then updates that model after each new data point. In this topic, you can build a workflow that suggests oxide compositions, ranks them, and refines the search as new thermal data comes in.

Why This Is a Good Topic

This is a strong science fair topic because it turns a huge materials search problem into a clear, testable algorithmic challenge. You can compare search strategies, measure how fast they find promising compositions, and test whether the model picks better candidates than random choice. The work connects to energy efficiency, waste heat recovery, and greener electronics. You can also learn real research skills like data cleaning, feature selection, model evaluation, and uncertainty analysis.

Research Questions

• How does Bayesian optimization compare with random search when selecting oxide compositions for low thermal conductivity? ?
• What is the effect of different material descriptors on the model’s ability to predict thermal conductivity? ?
• Does adding uncertainty estimates improve the next composition chosen by the active-learning loop? ?
• To what extent do high-entropy oxide composition ratios change predicted thermoelectric performance? ?
• Which acquisition function leads to the fastest improvement in candidate quality? ?
• How does training set size affect prediction error for thermal conductivity models? ?

Basic Materials

• Laptop or desktop computer with internet access.
• Python installed with scientific packages.
• Spreadsheet software for tracking compositions and results.
• Public materials dataset on oxides or thermoelectrics.
• Text editor or notebook for recording model settings and decisions.

Advanced Materials

• Access to a university materials database or literature corpus.
• Published thermal conductivity datasets for oxide materials.
• Materials informatics dataset with composition features and target properties.
• Access to high-performance computing or a strong workstation.
• Optional density functional theory outputs or simulated descriptors for candidate oxides.
• Reference software for materials analysis, such as pymatgen or matminer.

Software & Tools

• Python: Runs the modeling, active-learning loop, and data analysis.
• Jupyter Notebook: Lets you document each experiment and keep code, notes, and plots together.
• pandas: Organizes composition tables and cleans materials datasets.
• scikit-learn: Builds baseline regression models and checks prediction error.
• matplotlib: Plots learning curves, uncertainty, and candidate ranking.

Experiment Steps

1. Define the exact prediction target, such as thermal conductivity or a related proxy property.
2. Choose the composition space you will search, and set the limits so the problem stays manageable.
3. Build a baseline model first, so you can compare active learning against a simpler approach.
4. Select the feature set and acquisition function that will guide each new material suggestion.
5. Plan validation so you can test whether the model really improves after each new round of data.
6. Decide how you will report uncertainty, error, and search efficiency, not just final performance.

Common Pitfalls

• Using a dataset with mixed measurement methods, which makes thermal conductivity values hard to compare.
• Letting the model learn from too few examples, which can make the next suggested compositions look random.
• Feeding the algorithm poor composition descriptors, which hides the chemistry the model needs to see.
• Comparing active learning to a weak baseline, which makes the results less convincing.
• Treating predicted performance as proof, which ignores the gap between model output and real materials data.

What Makes This Competitive

A stronger project goes beyond building one model. You compare several acquisition functions, test multiple feature sets, and measure how quickly each method finds high-potential candidates. You can also explain where the model fails and why. If you connect prediction quality to materials chemistry, not just to code output, your project looks much more like real research.

Project Variations

• Use electrical conductivity or Seebeck coefficient instead of thermal conductivity as the target property.
• Compare high-entropy oxides with simpler binary or ternary oxides to test whether compositional complexity helps prediction.
• Add a literature-mining step that scores candidate oxides by both model uncertainty and publication evidence.

Learn More

• MIT OpenCourseWare Materials Science courses: Search MIT OpenCourseWare for materials science, thermodynamics, and computational materials content.
• NIST Materials Data Repository: Search the NIST data repository for materials datasets and examples of structured property data.
• Materials Project: Search the open materials database for crystal structures, computed properties, and composition data.
• PubMed: Search review articles on thermoelectric oxides, thermal conductivity, and materials informatics.
• Nature Reviews Materials: Read review articles on machine learning for materials discovery through your school library or public-access summaries.