AI Design of Anti-Perovskite Ionic Conductors

ISEF Category: Materials Science

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

The Hook

Battery materials do not have to be guessed one by one. A good model can search chemical space faster than a human can read it. Your project can test whether an AI system finds plausible anti-perovskite ionic conductors, then screen those ideas with physics-based rules. That mix of machine learning and materials chemistry makes this a strong research topic.

What Is It?

This project studies how a transformer model, a type of AI that learns patterns in sequences, can suggest new anti-perovskite materials. Anti-perovskites are crystal structures that can move ions well, which matters for batteries and solid electrolytes. Think of the model as a super-fast recipe recommender. Instead of food ingredients, it recommends combinations of atoms that might fit a crystal structure.

The Materials Project is a large public database of calculated materials properties. You can train or fine-tune a model on those structures, then ask it to propose new candidates. The bond-valence-sum filter acts like a sanity check. It estimates whether the ions in a proposed structure have reasonable bonding, so you can reject ideas that look random or unstable on paper.

Why This Is a Good Topic

This topic works well because you can measure clear outputs, such as how many valid candidates the model proposes, how often bond-valence screening rejects them, and how the suggestions compare with known materials. You also connect to a real problem, finding better solid electrolytes for safer energy storage. You can learn model evaluation, data cleaning, feature design, and materials screening without needing to synthesize anything in a lab.

Research Questions

• How does the choice of training set size affect the number of chemically valid anti-perovskite candidates the model generates?
• What is the effect of bond-valence-sum filtering on the fraction of proposed structures that remain plausible after screening?
• Does adding composition constraints change the diversity of generated anti-perovskite candidates?
• To what extent do generated candidates match known ionic conductor patterns in the Materials Project database?
• Which tokenization strategy produces the most valid crystal-structure suggestions?
• How does model temperature affect novelty versus chemical validity in generated materials?
• To what extent does a transformer outperform a simpler baseline model for proposing anti-perovskite compositions?

Basic Materials

• Laptop or desktop computer with at least 16 GB RAM.
• Python installed with Jupyter Notebook.
• Access to the Materials Project public database or downloaded open materials data.
• Spreadsheet software for tracking generated candidates and screening results.
• Text editor for cleaning composition files.
• Stable internet connection for database access and model documentation.

Advanced Materials

• University workstation or cloud GPU access for model training.
• Python environment with PyTorch or TensorFlow.
• Crystal structure files from the Materials Project or related open databases.
• Bond-valence-sum analysis tools or scripts.
• Materials informatics libraries such as pymatgen.
• Version control system such as Git for tracking model changes and experiments.

Software & Tools

• Python: Runs data cleaning, model training, and evaluation scripts for the project.
• Jupyter Notebook: Lets you explore data, test ideas, and document results in one place.
• pymatgen: Handles crystal structure files, composition parsing, and materials analysis.
• Matplotlib: Creates plots that compare validity, novelty, and screening results.
• Git: Tracks code versions and helps you compare model changes clearly.

Experiment Steps

1. Define the exact prediction task, such as generating new anti-perovskite compositions or structures.
2. Build a clean training set from open materials data and remove duplicates, missing values, and obvious outliers.
3. Choose a baseline method first, then compare it with your transformer model.
4. Design a screening pipeline that checks chemistry rules, structural plausibility, and bond-valence consistency.
5. Plan metrics for validity, novelty, diversity, and similarity to known ionic conductors.
6. Organize a comparison table so you can test whether the model creates useful candidates, not just many candidates.

Common Pitfalls

• Training on messy materials records, which makes the model learn inconsistent structure labels and noisy chemistry patterns.
• Treating every generated composition as a real candidate, which inflates results before bond-valence screening.
• Measuring only novelty, which can reward nonsensical formulas that no chemist would trust.
• Ignoring class imbalance in the database, which can push the model toward common materials and away from rare anti-perovskites.
• Comparing your model to no baseline, which makes it hard to tell whether the transformer actually adds value.

What Makes This Competitive

A stronger version of this project would not just generate random candidates. You would compare multiple model settings, use strict screening rules, and report several metrics at once, such as validity, novelty, diversity, and chemical plausibility. You could also test whether the model suggests compositions that known materials search methods miss. Careful error analysis and clear baseline comparisons would make the work much stronger.

Project Variations

• Try generating related solid electrolyte families instead of only anti-perovskites, then compare validity rates across structure types.
• Test whether a rules-based generator or a transformer produces more chemically plausible candidates from the same database.
• Add property filters for ionic conductivity proxies, then see how screening changes the tradeoff between novelty and likely performance.

Learn More

• Materials Project: Search the public database for crystal structures, compositions, and calculated properties related to solid electrolytes.
• Nature Materials: Search for review and research articles on anti-perovskites and ionic conductors through your school library or journal website.
• PubMed: Search review articles on solid electrolytes, crystal structure prediction, and materials informatics.
• NIH PubChem: Use it to check elemental and compound identifiers when you clean composition data.
• MIT OpenCourseWare: Search for materials science, machine learning, and computational chemistry course notes.