Teen Project: Build an AI-Powered Science Fair Display With Data and Explainability

Pick a question that can be tested (and measured)

An AI project is still a science fair project. The “AI” part is simply the tool you use to analyze patterns in data. The winning formula for many ai science fair projects for high school is:

• A real-world question your teen cares about
• A clear hypothesis
• Data they can collect or responsibly source
• A model that produces a prediction or classification
• An explanation a non-technical judge can understand

Here are a few strong, judge-friendly options (each can be done with a laptop and a phone):

• Can we predict classroom air quality from simple sensor readings? (temperature, humidity, CO₂ if available)
• Does screen time before bed predict sleep quality? (self-reported sleep score + phone usage stats)
• Can a model identify plant stress from leaf photos? (healthy vs. stressed leaves)
• What factors best predict recycling contamination in a household? (photos + checklist labels)

A solid machine learning science fair idea should have a target you can label. For example:

• Classification: “healthy vs. stressed leaf”
• Regression: “predict sleep score from habits”

Turn the idea into a hypothesis

Help your teen write a hypothesis that connects an input to an outcome:

• Hypothesis example (sleep): “If screen time within 60 minutes of bedtime increases, then my sleep quality score will decrease.”
• Hypothesis example (plants): “If a plant is water-stressed, leaf color and texture features will shift in a way a model can detect.”

A simple way to keep it scientific: your teen should identify:

• Independent variables (inputs): what you measure (screen time, humidity, image features)
• Dependent variable (output/label): what you’re predicting (sleep score, stressed vs. healthy)
• Controls: what you try to keep consistent (same room, same lighting for photos)

Collect data you can defend (and analyze)

Most science fair AI projects fail at one point: the data is too small, messy, biased, or unlabeled. The goal isn’t “big data.” It’s trustworthy data.

A realistic data plan for 2–3 weeks

Aim for:

• 50–200 rows for spreadsheet-style projects (sleep, recycling, air quality)
• 200–600 images for photo projects (plants, trash sorting, handwritten notes)

If your teen can’t collect that much, they can still succeed—just be honest about limitations and use simpler models.

What to track every day

A strong science fair project with data analysis includes a “data dictionary” (a mini legend for what each column means). Here’s a practical template your teen can copy into a notebook or Google Sheet.

Feature (column)	What it means	How to collect	Type	Quality tip
date	Day of observation	Calendar	ID	Use consistent format (YYYY-MM-DD)
screen_time_60min	Minutes of phone use in last 60 min before bed	Phone stats	Numeric	Log it at the same time nightly
bedtime	Time you attempted sleep	Self log	Numeric/time	Convert to minutes past midnight
caffeine_after_2pm	Did you have caffeine after 2pm?	Yes/No	Binary	Define “caffeine” upfront
exercise_minutes	Minutes of exercise	Self/fitness app	Numeric	Include “walking” rules
sleep_score	Morning rating 1–10	Self log	Label	Write a scoring rubric

This table does two things judges love:

• Shows your teen planned measurements carefully
• Makes the project reproducible (someone else could repeat it)

Quick data checks (simple, but powerful)

Before training any model, your teen should do basic checks and record what they find:

• Missing values: Any blank days? Why?
• Outliers: One day with 400 minutes of screen time—was it real?
• Balance (for classification): If 90% of images are “healthy,” the model can “cheat.”
• Train/test split: Save a portion for final testing (don’t peek too early).

Parent tip: encourage a “lab notebook” mindset. Judges often reward careful documentation as much as fancy code.

Build the model (keep it simple, make it measurable)

The most impressive high school displays usually use simple models well explained, not complicated models poorly explained.

Choose a model that matches the data

• Spreadsheet data (sleep, air quality): start with linear regression (predict a number) or logistic regression (predict yes/no).
• Image data (plants, recycling photos): try a pre-trained image classifier (transfer learning). This uses a model that already learned general visual features.

If your teen is coding, common tools include:

• Python + scikit-learn (tabular data)
• Python + Teachable Machine / TensorFlow (images)

What matters for the science fair: define the pipeline clearly.

Define your success metrics

This is where many students lose points: they say “it worked,” but can’t quantify “worked.” Pick one or two metrics and stick to them.

• For classification (healthy vs. stressed):
  • Accuracy (overall correct)
  • Precision/recall (especially if one class is rarer)
  • Confusion matrix (easy to show on a poster)
• For regression (sleep score):
  • Mean Absolute Error (average points off)
  • R² (how much variance is explained)

A science-fair-friendly experiment design

To show this is a science project, not just “I trained a model,” your teen should compare at least two conditions:

• Baseline vs. model:
  • Baseline: “predict the average sleep score every day”
  • Model: regression with features
• Small vs. larger dataset:
  • Train on Week 1 only vs. Weeks 1–3
• Different feature sets:
  • Model A uses screen time + bedtime
  • Model B adds caffeine + exercise

Even if the performance gain is small, the comparison is what tells a research story.

Explainability: how to show what the AI learned

Judges (and parents) often ask: “Okay… but why did the model decide that?” This is where explainability turns a good project into a great one.

If your teen is wondering how to explain ai model in project, give them this simple framework:

• What went in? (inputs/features)
• What came out? (prediction)
• What mattered most? (top features or visual regions)
• How sure was it? (confidence)
• When does it fail? (limitations)

Explainability for spreadsheet-style models

For linear/logistic regression, explanation can be very concrete:

• Feature importance (weights): “Screen time had the strongest negative effect.”
• Partial dependence (simple version): show a small chart: more screen time → lower predicted sleep.

What to put on the board:

• A bar chart of the top 3–5 most important features
• 2–3 example predictions with the input values listed

Explainability for image models

Image models can feel like a “black box,” so focus on interpretability tools that are easy to explain:

• Confidence scores: “The model predicted ‘stressed’ with 82% confidence.”
• Example gallery: show 5 correct + 5 incorrect predictions.
• Heatmaps (optional): If your teen can generate a simple attention/activation map (like Grad-CAM), it’s powerful—just keep the explanation plain.

A judge-friendly line your teen can use:

• “The model looks at patterns of color and texture. In stressed leaves, it often focuses on yellowing edges and spotted regions.”

Don’t hide limitations—feature them

A mature project clearly states failure modes:

• Lighting changes confuse the leaf model
• Self-reported sleep scores are subjective
• Small dataset limits generalization

Judges tend to trust students who understand uncertainty.

Next Steps: build a science fair display that tells the whole story

A winning AI board isn’t a wall of code. It’s a clear story: hypothesis → data → model → results → explanation.

Here’s a simple layout your teen can follow:

• Panel 1: Question + Hypothesis
  • One sentence question
  • One sentence hypothesis
  • Why it matters
• Panel 2: Data Collection + Controls
  • Data table sample (5–10 rows)
  • How labeling worked
  • What you controlled
• Panel 3: Model + Metrics
  • Diagram of the pipeline (collect → clean → train → test)
  • Metrics + confusion matrix or error chart
  • Baseline comparison
• Panel 4: Explainability + Examples
  • Feature importance or example predictions
  • 3 “model got it right” + 3 “model got it wrong”
  • Limitations and next improvements

A practical 10-day plan (you can actually fit into a busy schedule)

• Days 1–2: Pick question, define label, write hypothesis, set up data sheet
• Days 3–6: Collect data daily, do quick cleaning checks, label carefully
• Days 7–8: Train baseline + model, record metrics, save screenshots/plots
• Day 9: Create explainability visuals (feature bars, example gallery)
• Day 10: Assemble board, practice a 60-second explanation + a 3-minute deep dive

If you want a fast “parent checklist”

• Your teen can explain the dataset in 30 seconds
• They have a baseline to compare against
• They can show at least one metric (not just “it worked”)
• They can describe one limitation and one improvement

If you’d like, Intellect Council lessons can help your teen go from data collection to a working model with guided steps—without turning the project into a confusing college-level assignment. The goal is confidence, clarity, and a project they’re proud to present.