Persistent Homology for MOF Gas Separation

ISEF Category: Materials Science

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

The Hook

Some materials can trap one gas and let another pass through. That trick powers carbon capture, hydrogen storage, and cleaner chemical processing. The weird part is that a material’s pore shape can matter as much as its chemistry. You can study that shape with topology, the math of holes, loops, and connected spaces.

What Is It?

Metal-organic frameworks, or MOFs, are crystalline materials made from metal nodes and organic linkers. Think of them like 3D scaffolds with tiny tunnels and chambers. Those pores control which gas molecules can fit, stick, or slip through.

Persistent homology is a way to measure shape across many length scales. Instead of asking, “How many pores do you see?”, you ask, “Which voids, channels, and loops stay present as I change the measurement scale?” That gives you a fingerprint of the pore network. You can then compare that fingerprint with gas-separation selectivity, which tells you how well a MOF favors one gas over another.

This project lives at the intersection of materials science, math, and data science. You are not just describing a structure. You are testing whether topological features can predict a material property that matters in the real world.

Why This Is a Good Topic

This is a strong science fair topic because you can test a clear question with public data. The CoRE-MOF database gives you many materials to compare, and persistent homology gives you a new way to turn structure into numbers. That makes the project measurable, computational, and open-ended. You can also learn real research skills, like feature engineering, model testing, and performance comparison.

Research Questions

• How does persistent homology of MOF pore networks relate to gas-separation selectivity?
• What is the effect of using different topological feature sets on prediction accuracy for selectivity?
• Does a model built from persistent homology outperform one built from simple geometric descriptors?
• To what extent do pore-network features predict selectivity for one gas pair better than another?
• Which persistence summary, such as Betti curves or persistence images, gives the best results for MOF selectivity prediction?
• How does the number of MOFs in the training set affect the stability of a topology-based model?
• What is the effect of removing outlier MOFs on the relationship between pore topology and selectivity?

Basic Materials

• Computer with a modern processor and enough memory for data analysis programs.
• Stable internet access for downloading MOF structure data and property tables.
• Spreadsheet software for cleaning and organizing sample metadata.
• Python installed with scientific libraries for analysis.
• Text editor or notebook environment for writing code and notes.
• Access to the CoRE-MOF database through a published source or mirrored dataset description.
• Cloud storage or external drive for backing up files.

Advanced Materials

• Computer with strong memory and CPU or access to a university cluster.
• Python environment with topology and machine learning libraries.
• MOF structure files from the CoRE-MOF database.
• Gas-separation selectivity labels from a curated literature or database source.
• Graph or point-cloud processing tools for converting pore geometry into analysis inputs.
• Version control system for tracking code changes and analysis runs.
• Optional access to crystal visualization software for inspecting structures.

Software & Tools

• Python: Handles data cleaning, feature extraction, model building, and statistics for the project.
• Jupyter Notebook: Keeps code, plots, and notes together in one place.
• scikit-learn: Builds and tests machine learning models for prediction tasks.
• GUDHI: Computes persistent homology features from geometric data.
• Ripser: Calculates persistent homology quickly for point-cloud style inputs.
• matplotlib: Makes clear plots for comparing model performance and feature patterns.

Experiment Steps

1. Define the exact gas-separation property you want to predict and choose one narrow gas pair or target metric.
2. Select a clean MOF subset from the database, then decide how you will handle missing labels, duplicates, and extreme outliers.
3. Convert each MOF structure into a pore-network representation that your topology method can read.
4. Choose how you will summarize persistence output into features that a model can use.
5. Build a baseline model with simple descriptors, then compare it against your topology-based model.
6. Plan evaluation rules that test whether your result holds up across different train-test splits and performance metrics.

Common Pitfalls

• Mixing MOFs with different property labels, which makes the model learn inconsistent targets.
• Feeding raw crystal files into persistent homology without a clean pore representation, which turns the topology result into noise.
• Using too many features for a small dataset, which causes the model to memorize the training set instead of learning a pattern.
• Comparing models with different data splits, which makes the performance numbers hard to trust.
• Ignoring simple baseline descriptors, which makes it impossible to tell whether topology adds real predictive value.

What Makes This Competitive

A strong version of this project does more than run one model. You compare topology features against solid baselines, test several persistence summaries, and explain when the method works or fails. You can also separate gas pairs, pore size ranges, or MOF families to see where the signal changes. That kind of careful analysis shows real judgment, not just code execution.

Project Variations

• Study whether persistent homology predicts CO2 over N2 selectivity better than it predicts H2 over CH4 selectivity.
• Compare pore-topology features from MOFs with high surface area against MOFs with smaller, more selective pores.
• Test whether a graph-based pore representation or a voxel-based pore representation gives stronger persistence features for selectivity prediction.

Learn More

• MIT OpenCourseWare: Search for free materials on topology, data analysis, and machine learning foundations.
• PubChem: Use it to check gas properties and molecule identifiers when you compare adsorption targets.
• CoRE-MOF database papers in Nature Communications or related journals: Search journal articles for curated MOF datasets and how researchers filter structures.
• GUDHI documentation: Read the free library guide for persistent homology and persistence images.
• scikit-learn user guide: Find model selection, cross-validation, and regression tools for comparing feature sets.
• NIH PubMed: Search for review articles on metal-organic frameworks, gas separation, and topological data analysis.