Predicting Biogas Yield With Machine Learning

ISEF Category: Energy: Sustainable Materials and Design

Subcategory: Biological Process and Design  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Food waste can turn into fuel, but not all waste makes the same amount. A banana peel and a dairy slurry do not behave like the same input. Your job is to teach a model to spot those differences before digestion even starts. That gives you a clean, data-driven project with real energy stakes.

What Is It?

Biogas is a gas mix, mostly methane and carbon dioxide, made when microbes break down organic material without oxygen. Think of it like a sealed stomach for waste, where the recipe of the input changes the gas output. Feedstock composition means what the waste contains, like fiber, fat, protein, moisture, carbon, and nitrogen.

In this project, you use public USDA waste-stream or food-waste datasets, plus any biogas yield data you can find in public sources, to train a machine-learning model. Machine learning means a program looks for patterns in data and makes predictions from them. Your model tries to predict biogas yield from the input traits, then you test how well it works on data it has never seen.

Why This Is a Good Topic

This makes a strong science fair topic because you can ask a real prediction question and test it with public data. You do not need a wet lab to get started, and you can still do serious analysis with cleaning, feature selection, model comparison, and validation. The topic connects to renewable energy, food waste reduction, and landfill diversion, so the real-world value is easy to explain. You can also scale the project up if you find better datasets or more careful ways to measure model error.

Research Questions

• How does feedstock carbon-to-nitrogen ratio affect predicted biogas yield?
• What is the effect of moisture content on model accuracy when predicting biogas yield?
• Does adding nutrient composition features improve prediction more than using mass-based waste categories alone?
• To what extent can a random forest model outperform linear regression on public biogas data?
• Which feedstock variables contribute most to predicted methane yield across USDA waste streams?
• How does combining food waste categories change predicted yield compared with modeling each category separately?

Basic Materials

• Laptop or desktop computer with internet access.
• Spreadsheet software such as Google Sheets or Excel.
• Python installed with Jupyter Notebook or access to Google Colab.
• Public USDA waste-stream dataset or related public biomass composition dataset.
• Public biogas yield dataset from a government, university, or peer-reviewed source.
• Notes document for tracking data sources, variables, and cleaning choices.
• Basic citation manager or reference list file.

Advanced Materials

• High-performance laptop or desktop computer.
• Python in Jupyter Notebook, with pandas, scikit-learn, matplotlib, and seaborn.
• Access to a second public dataset for external validation.
• Statistical software such as R or Python stats libraries for residual analysis.
• Feature engineering scripts for ratios, interaction terms, and encoded categories.
• Git or another version-control tool for tracking code changes.
• Optional cloud notebook access for larger datasets.

Software & Tools

• Python: Lets you clean data, build models, and test prediction accuracy.
• Google Colab: Runs Python notebooks in the browser without local setup.
• Jupyter Notebook: Keeps code, outputs, and notes in one place for analysis.
• scikit-learn: Provides regression models, train-test splitting, and model evaluation tools.
• PubMed: Helps you find review articles and studies on anaerobic digestion and biogas yield.

Experiment Steps

1. Define the prediction target and decide whether you will model total biogas, methane yield, or both.
2. Select a public dataset that links feedstock composition to measured gas output, then record which variables are complete enough to use.
3. Clean the dataset, standardize units, and decide how you will handle missing values and outliers.
4. Build a simple baseline model first, then compare it with at least one nonlinear model.
5. Split the data into training and test sets, then choose evaluation metrics that match your goal.
6. Check which features matter most, then test whether those patterns still hold on a separate dataset or sample group.

Common Pitfalls

• Mixing datasets that define biogas yield in different units, which makes the target variable inconsistent.
• Using feedstock categories with too many missing composition values, which weakens the model before training starts.
• Letting one dominant waste type overpower the rest of the data, which makes the model look better than it really is.
• Skipping a simple baseline, which makes it hard to prove that machine learning adds value.
• Treating correlation as cause, which leads to claims about chemistry that the data cannot support.

What Makes This Competitive

A stronger version of this project goes beyond one model and one dataset. You can compare several model types, test them on outside data, and explain why some feedstock features matter more than others. You can also look for bias, missing-data patterns, and performance gaps across waste categories. That kind of careful analysis shows judgment, not just coding.

Project Variations

• Use only food-waste categories and test whether kitchen waste predicts biogas better than mixed municipal organic waste.
• Compare linear regression, random forest, and gradient boosting for methane yield prediction from composition features.
• Add a climate or logistics angle by asking how source, transport distance, or season changes the usable feedstock profile.

Learn More

• USDA Economic Research Service: Search for reports and datasets on food waste, biomass, and agricultural residues.
• NOAA National Centers for Environmental Information: Find climate and weather data if you want to test seasonal effects on waste streams.
• NIH PubMed: Search review articles on anaerobic digestion, biogas yield, and feedstock composition.
• MIT OpenCourseWare: Look for free course materials on data science, statistics, and environmental engineering topics.
• Bioresource Technology: Search the journal for peer-reviewed studies on biogas yield prediction and feedstock analysis.