PEWP: Predicting Exoplanet Weather Patterns

Using machine learning to model exoplanet atmospheres from Earth and Mars climate data.

Comment

Inspiration

The first time we scrolled through raw Mars InSight weather logs, the numbers felt almost alien. Temperatures swung from daytime highs barely tolerable by Antarctic standards to nighttime lows that would freeze steel. The pressure column looked more like a vacuum gauge than an atmosphere. Compare that with Earth’s steady rhythms, and you see two planets that couldn’t feel more different: yet both recorded in neat rows of data.

That contrast planted a question: if we can learn the language of weather from Earth and Mars, can we translate it to worlds light-years away? Exoplanets rarely give us more than a radius, a mass, and an orbit. Still, maybe those fragments are enough to sketch their skies.

What it does

PEWP is a machine learning framework that forecasts temperature, pressure, and wind on exoplanets. It uses atmospheric data from Earth and Mars, then extrapolates to planets where only a few constants are known. By training on real, detailed data, the model can make informed predictions about environments far outside our reach.

How we built it

We combined multiple open datasets:

  • NASA Exoplanet Archive for planetary parameters.
  • NASA InSight for Martian atmospheric records.
  • Copernicus Climate Data Store for Earth climate variables.

These were cleaned, resampled, and scaled to produce multiple consistent datasets across planets. We then trained a Random Forest Regressor to predict atmospheric conditions.

Challenges we ran into

The greatest challenge, and the most time-consuming was simply finding the data as there is very limited information available about exoplanets. Even when the data we needed existed, it was often part of enormous datasets that were hard to sort. For example, the Mars InSight Data source we originally found gave a 30 MB spreadsheet, giving minute-by-minute data, which made it impossible to use that source effectively. Additionally, calling the API's and accessing downloads for certain datasets took upwards of 30 minutes. We waited for nearly an hour for one of our data sets, but once we went to download it, we discovered it was a 300 MB grib file which was not usable on our computers.

Accomplishments that we're proud of

We built a complete pipeline that transforms raw planetary data into reliable atmospheric predictions. The model performed with high accuracy (R^2 value of 0.9421), and more importantly, it demonstrated that two well-characterized atmospheres can teach us enough to make informed forecasts about exoplanets. Turning scattered and inconsistent data into a functioning predictive system was itself a significant achievement.

What we learned

The project reinforced how critical preprocessing and data engineering are in planetary science. We learned how to comb through massive, irregular datasets and reduce them without losing meaning, and how to adapt machine learning when data is scarce or incomplete. We also saw that environmental models can remain powerful even when training examples are extremely limited, as long as they capture meaningful variation.

What's next for PEWP: Predicting Exoplanet Weather Patterns

Expand training data with more planetary atmospheres. Venus, for example, would work well as an archetype for exoplanets with very thick atmospheres, while Europa could serve as a model for exoplanets with very thin atmospheres. These changes would give our model greater range and accuracy when dealing with more extreme environments.

Bibliography

Built With

  • caltech
  • copernicus
  • era5
  • mars-weather-service
  • nasa
  • nasa-api
  • python
  • skribb
Share this project:

Updates