Climate science once belonged to research labs with super-computers. Not anymore. During our spring HigherEds AI Bootcamp, four high-school juniors built a working climate-trend predictor on a stack no fancier than Google Colab, public NOAA data and the open-source TensorFlow library. Their goal: forecast regional temperature anomalies five years out and visualise the uncertainty so policy clubs in their schools could debate mitigation plans. What follows is a play-by-play of how they moved from raw CSV files to an interactive dashboard; proof that motivated teenagers, the right mentorship and good tooling can tackle global challenges.
From Idea to Dataset
The team started with a pain point they could relate to. Two members serve on their city’s Youth Sustainability Council; they kept hearing statements like “winters feel warmer” without numbers to back them up. After a planning session with their bootcamp mentor, they chose NOAA’s Global Historical Climatology Network because it offers daily land-surface temperatures dating back to 1880.
Data volume: 4.8 GB
Stations analysed: 7 937
Period of interest: 1980-2023 (to align with satellite verification)
Each student owned one phase of the extract-transform-load pipeline. Python scripts merged station records, imputed missing values with linear interpolation and stored the result in a single Parquet file for fast reading.
Building the Model
Rather than jump straight into deep learning, the cohort benchmarked three baselines:
Model | MAE (°C, validation set) | Notes |
|---|---|---|
Linear regression | 0.72 | Rapid to train; under-fitted annual cycles |
Random forest | 0.58 | Captured non-linearity; lagged on multi-year drift |
LSTM (TensorFlow) | 0.41 | Best seasonal + trend tracking |
The LSTM architecture used two recurrent layers of 128 units, dropout of 0.2, and Adam optimiser at 1e-3. Training ran for 30 epochs on a free Tesla T4 GPU in Colab, taking just under 11 minutes. Students monitored loss curves in real time and practised early-stopping logic to avoid over-fitting, exactly the skill set employers cite when screening junior data-science résumés.
Visualising Uncertainty
Forecasts are only as useful as their confidence bands. The team wrapped the model in a small Streamlit app that plots predicted temperature anomaly +/- one standard deviation. Hovering over any point reveals the numeric range so debate-team users can quote figures accurately. One mentor challenge was to create a “what-if” slider that adjusts CO₂ concentration scenarios. By feeding IPCC pathway multipliers into the trained network, students generated side-by-side curves: business-as-usual versus aggressive emission cuts.
Impact Beyond the Classroom
Science-fair acclaim: The project won Best Data Innovation at the State Enviro-Tech Expo.
Policy influence: A summary dashboard now sits on the city council’s intranet; staff use it during monthly climate-action briefings.
Scholarship leverage: Two team members cited the project in college essays and secured merit awards from environmental foundations.
A senior data-analyst from an energy firm, invited as an external judge, summarised the significance: “They delivered a defensible model, but more impressive is their understanding of uncertainty. That sophistication is rare even in entry-level hires.”
Take-Home Lessons for Future Bootcampers
Start simple, then iterate: Baselines set expectations and teach humility.
Own the data pipeline: Cleaning and feature engineering consumed 60 percent of project hours; shortcuts there sink forecasts later.
Explain, don’t just predict: Visualising variance turned raw numbers into actionable insight.
Collaborate like a squad: Shared GitHub board, documented commits and weekly retros kept four teens aligned without drama.
If a group of high-schoolers can demystify climate forecasting in nine weeks, imagine what your teen could build with structured guidance and expert tutoring. Consider enrolling them in our next AI Bootcamp cohort and turn curiosity into impact.