Neural Network Climate Sensitivity Modeling

ISEF Category: Earth and Environmental Sciences

Subcategory: Atmospheric Science  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A small change in greenhouse gases can shift Earth’s temperature by degrees, not fractions of a degree. That is why climate models matter. You can build a neural network that learns the same physics climate scientists study, then test how humidity changes the answer. This project puts machine learning and atmospheric science on the same team.

What Is It?

This project uses a physics-informed neural network, or PINN. That is a neural network that does not just learn from examples, it also has to obey equations from physics. In this case, the model learns a radiative-convective equilibrium temperature profile, which is a fancy way of saying it predicts how temperature changes with height when radiation and rising air balance out.

Think of the atmosphere like a layered blanket. Sunlight warms it from above and below, while heat also escapes back to space. A PINN tries to learn where that balance lands. You then change inputs like CO₂ and humidity and ask how much the surface temperature would change if CO₂ doubled.

Why This Is a Good Topic

This is a strong science fair topic because you can ask clear, measurable questions and test them with real atmospheric equations. It connects to a real problem, climate sensitivity, which matters for weather planning, policy, and long-term risk. You can learn machine learning, numerical modeling, and scientific analysis in one project, and the final result can be compared against published climate ranges.

Research Questions

• How does assumed relative humidity change the predicted surface warming from CO₂ doubling?
• What is the effect of different atmospheric layer counts on the stability of the learned temperature profile?
• Does adding a physics loss term improve agreement with known radiative-convective equilibrium behavior?
• To what extent do training data choices change the model's climate sensitivity estimate?
• Which humidity profile gives the largest change in predicted lapse rate, the rate temperature drops with height?
• How does the network's predicted temperature profile compare with a simple baseline radiative-convective model?
• To what extent does CO₂ forcing change the altitude of the tropopause, the boundary between lower and upper atmosphere?

Basic Materials

• Laptop or desktop computer with a modern GPU or access to cloud computing
• Python installed with NumPy, Pandas, Matplotlib, and a machine learning library such as PyTorch or TensorFlow
• Access to atmospheric physics equations from a textbook or university notes
• Published climate or atmospheric profile data from NASA or NOAA
• Spreadsheet software for tracking experiments and results
• Notebook for design notes and model checks.

Advanced Materials

• University or cloud compute access with a dedicated GPU
• Python with PyTorch or TensorFlow, SciPy, and xarray
• Atmospheric profile datasets from reanalysis sources such as NOAA or NASA
• Reference radiative transfer or climate model outputs for validation
• Version control with Git for experiment tracking
• Jupyter Notebook for code, figures, and analysis.

Software & Tools

• Python: Builds and trains the physics-informed neural network and runs the climate sensitivity experiments.
• Jupyter Notebook: Keeps code, plots, and notes in one place for easy review.
• PyTorch: Supports custom neural network training with physics-based loss terms.
• Matplotlib: Makes temperature profile and sensitivity plots.
• Git: Tracks model changes, results, and failed attempts.

Experiment Steps

1. Define the climate question you want your model to answer, then pick one output to predict first, such as surface temperature or the full temperature profile.
2. Translate the atmospheric physics into equations your network must satisfy, then decide which terms belong in the physics loss.
3. Choose a baseline model and a PINN version so you can compare learning with and without physical constraints.
4. Plan your input features, output targets, and validation method before training, so you can test whether humidity assumptions change the result.
5. Design a comparison set for CO₂ doubling, then decide how you will measure climate sensitivity from the model outputs.
6. Set up plots and statistical checks that let you compare predicted profiles, error patterns, and sensitivity ranges across assumptions.

Common Pitfalls

• Training a neural network on too little atmospheric data, which makes the temperature profile look smooth but physically wrong.
• Changing the humidity assumption and the training setup at the same time, which makes it impossible to tell what caused the result.
• Using a physics loss that is too weak, which lets the model ignore radiative-convective balance.
• Comparing outputs without matching the same boundary conditions, which creates fake differences between models.
• Reporting one climate sensitivity number without uncertainty bounds, which hides how unstable the prediction is.

What Makes This Competitive

A competitive version of this project goes beyond making a model that runs. You would compare multiple humidity parameterizations, test several physics loss weights, and show how each choice changes climate sensitivity. Strong entries also check uncertainty, not just average error. If you can explain why one modeling choice shifts the answer more than another, you are doing real scientific analysis.

Project Variations

• Use a simpler radiative-convective model as the baseline and compare it against the PINN on the same atmosphere profile data.
• Test how the predicted climate sensitivity changes when you switch from fixed relative humidity to height-dependent humidity.
• Extend the model to compare Earth-like atmospheres with slightly different solar input or surface pressure assumptions.

Learn More

• NASA Earthdata: Search for atmospheric and climate datasets, reanalysis products, and background material on Earth’s atmosphere.
• NOAA National Centers for Environmental Information: Find climate observations, atmospheric records, and educational resources on environmental data.
• MIT OpenCourseWare, Atmospheric Science courses: Search course notes and lectures on radiative transfer, convection, and climate basics.
• Atmospheric Chemistry and Physics: Search the journal for review articles on radiative-convective equilibrium and climate sensitivity.
• PubMed: Search for review articles on machine learning in climate science and physics-informed neural networks.
• NASA Global Climate Change: Find clear summaries of greenhouse forcing, CO₂, and climate sensitivity from a trusted source.