ML Inverse Design of Biodegradable Polymer Blends

ISEF Category: Biomedical Engineering

Subcategory: Biomaterials and Regenerative Medicine  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Biodegradable implants need to disappear on a schedule. Too fast, the wound reopens. Too slow, the patient needs a second surgery to remove fragments. By scraping degradation rates from published papers, you can train a small ML model that picks PLA-PCL-PHA ratios for any target half-life. Then a backyard compost pile becomes your test bench.

What Is It?

PLA, PCL, and PHA are biodegradable polyesters used in resorbable medical devices. Each degrades at a different rate. Blends behave non-linearly, so picking the right ratio is hard.

Literature mining means extracting structured data (composition, degradation half-life, environment) from published abstracts and tables. A small dataset of about 100 to 300 rows is enough to train a regression model that maps composition to half-life.

Compost is a simplified, accessible analog to simulated body fluid. Microbial activity, moisture, and temperature drive ester-bond hydrolysis. Weighing buried film samples over weeks gives a degradation curve you can compare to the model's prediction.

Why This Is a Good Topic

Inverse design with ML is highly ranked at ISEF and combines chemistry, biology, and computer science. The data is free, the model fits on a laptop, and the validation step is real-world. You will learn dataset curation, supervised regression, and field-style validation.

Research Questions

• How does PLA fraction predict compost half-life across the literature?
• What is the effect of crystallinity feature inclusion on model error?
• Does the model trained on SBF data extrapolate to compost data?
• To what extent does sample thickness shift validation half-life?
• Which blend matches a 4-week target half-life with smallest error?
• How does compost moisture affect measured degradation?
• What is the effect of test temperature on predicted-vs-measured agreement?

Basic Materials

• PLA, PCL, and PHA pellets or sheets (maker suppliers).
• Hot plate and aluminum molds.
• Compost pile or home compost bin.
• Mesh bags for sample retrieval.
• Digital scale (0.001 g accuracy).
• Soil thermometer and moisture meter.
• Laptop with Python.

Advanced Materials

• Differential scanning calorimeter.
• Gel permeation chromatography.
• Lab-grade PCL and PHA of known molecular weight.
• Environmental chamber.

Software & Tools

• Python (scikit-learn): Trains regression models on the blend dataset.
• Pandas: Cleans and joins literature data tables.
• Matplotlib: Plots model predictions vs. measurements.
• Zotero: Tracks citations for every data row.

Experiment Steps

1. Decide the inclusion criteria for literature data (same SBF, same temperature, reported thickness) and document them.
2. Build a feature set (composition, thickness, environment) before any modeling.
3. Split the dataset by paper, not by row, so leakage cannot inflate accuracy.
4. Choose three predicted compositions to fabricate and validate.
5. Plan controls (pure PLA, pure PCL) in the compost test.
6. Compare measured half-lives to predicted intervals and discuss disagreements.

Common Pitfalls

• Mixing degradation data from SBF and compost without flagging the environment.
• Reporting test accuracy after leaking rows from the same paper into training.
• Buried samples drying out, halting hydrolysis without notice.
• Casting blend films with uneven thickness and skewing degradation.
• Reading the literature half-life from a fitted curve, not the raw datapoint.

What Makes This Competitive

Build the dataset with citations and version it. Compare at least two model classes (random forest and gradient boosting) and report cross-validated error. Pick three predicted-best blends, validate them in compost, and report whether the model's predicted half-life lies inside the measured confidence interval. Add an ablation study removing one feature at a time.

Project Variations

• Replace random forest with a small neural network and compare interpretability.
• Add temperature as a feature and predict tissue-temperature half-life.
• Use a Bayesian model to report calibrated uncertainty on each prediction.

Learn More

• PubMed: Search PLA PCL PHA blend degradation reviews.
• NIH PubMed Central: Open-access datasets on resorbable polymers.
• scikit-learn documentation: Free regression tutorials.
• MIT OpenCourseWare: Course 6.036 Introduction to Machine Learning.
• USDA composting research publications: Free home composting standards.