Train a Vision Model for Reaction Outcomes

ISEF Category: Chemistry

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: 1 to 2 Months

The Hook

A phone camera can see a reaction before your eyes can name it. That matters in a classroom where students may miss weak color shifts, tiny precipitates, or brief bubbling. You can turn those messy video clips into data a model can learn from. That makes chemistry more measurable, and more accessible.

What Is It?

This project uses machine learning, which means you train a computer to spot patterns from examples. In this case, the examples are short videos of chemistry reactions. The model learns to tell whether a reaction made a precipitate, changed color, produced gas, or did more than one of those things.

Think of it like teaching a friend to sort photos. You show enough labeled examples, and the friend starts noticing clues, like cloudiness for a precipitate or rising bubbles for gas. Your job is to decide what the model should look at, how you will label each clip, and how well it performs on new videos it has never seen before.

The chemistry part matters too. Reaction outcomes are not random. They depend on the substances, concentrations, lighting, camera angle, and background. Your project can test how well a vision model handles those real classroom conditions.

Why This Is a Good Topic

This is a strong science fair topic because you can test a clear question, measure accuracy, and compare design choices. You do not need a fancy lab to collect videos, but you do need careful labels and clean experimental design. The real-world value is easy to explain, since under-resourced schools often need tools that help students record and interpret reactions more consistently. You can learn both chemistry and basic ML evaluation without needing advanced theory first.

Research Questions

• How does video lighting affect a model’s ability to classify precipitate, color change, and gas evolution?
• What is the effect of training set size on reaction outcome classification accuracy?
• Does adding background subtraction improve classification of faint precipitation events?
• To what extent do different camera angles change model performance on the same reaction?
• Which feature set, raw frames or motion-based features, gives better outcome classification?
• How does class imbalance affect precision for rare reaction outcomes?
• What is the effect of using student-recorded videos versus standardized lab videos on generalization?

Basic Materials

• Smartphone or tablet camera with video recording
• Tripod or stable phone stand
• Clear reaction containers or test tubes
• Common safe chemistry reagents approved by your teacher
• White poster board or plain backdrop
• Consistent desk lamp or ring light
• Spreadsheet software for labeling data
• Computer with internet access
• Notebook for recording reaction labels and conditions.

Advanced Materials

• School lab glassware and reaction supplies approved by your teacher
• Several smartphones or cameras for varied viewpoints
• Light meter or exposure app
• Color calibration card
• Computer with GPU access if available
• Python environment with ML libraries
• ImageJ for frame inspection and color checks
• Secure storage for video datasets and labels.

Software & Tools

• Python: Runs your data cleaning, model training, and evaluation scripts.
• Google Colab: Gives you a free cloud notebook for training simple vision models.
• ImageJ: Helps you inspect frames, crop regions, and compare color changes.
• OpenCV: Extracts video frames and basic motion features from reaction clips.
• scikit-learn: Builds baseline classifiers and scores accuracy, precision, and recall.

Experiment Steps

1. Define the exact reaction outcomes you will label, and decide whether one clip can have more than one label.
2. Choose one visual input type to start with, such as full frames, cropped reaction areas, or motion summaries.
3. Build a labeling rule that different people could apply the same way to the same video.
4. Split your videos into training, validation, and test groups before you train anything.
5. Compare at least one simple baseline model with one vision model so you know whether the AI adds value.
6. Plan a stress test for lighting, background, or camera angle to see where the model breaks first.

Common Pitfalls

• Using clips with mixed outcomes, which makes the model learn blurry labels instead of clear classes.
• Recording reactions under changing light, which causes the same color change to look different across sessions.
• Cropping the reaction area too tightly, which can cut off bubbles, cloudiness, or other useful motion cues.
• Mixing training and test clips from the same reaction batch, which inflates accuracy and hides weak generalization.
• Ignoring class imbalance, which can make the model look accurate while missing rare gas or precipitate events.

What Makes This Competitive

A competitive version of this project goes beyond a simple classifier. You compare multiple model inputs, test how well the model generalizes to new lighting or new students, and report the failure modes clearly. Strong entries often add careful labeling rules, confusion matrices, and statistical checks instead of stopping at one accuracy score. You can also make the project stronger by asking which reaction features matter most, not just whether the model works.

Project Variations

• Use only household-safe reactions and test whether the model still separates color change from precipitate events.
• Compare smartphone video against microscope-style close-up video for the same outcome labels.
• Train the model on one set of classroom videos, then test whether it generalizes to a different teacher’s lab setup.

Learn More

• MIT OpenCourseWare, Introduction to Machine Learning: Search MIT OpenCourseWare for a free course that covers model evaluation and basic supervised learning.
• scikit-learn User Guide: Read the official docs for simple classifiers, train-test splits, and confusion matrices.
• OpenCV Documentation: Learn how to extract frames, crop regions, and measure motion in video clips.
• PubMed: Search for review articles on computer vision in chemistry education and automated reaction monitoring.
• NIH ImageJ: Find tutorials and docs for measuring color and image intensity in scientific images.
• Nature Communications and Journal of Chemical Education: Search these journals for papers on reaction monitoring, chemistry education, and vision-based analysis.