Skip to content

How it works

From NASA's ISS flame tests to the next test to run.

The four steps

  1. Step 1: Collect ISS tests

    NASA burned materials aboard the ISS.

  2. Step 2: Check each cabin

    Did those tests match the cabin's air?

  3. Step 3: Train the model

    It learns when flames burn, and says when it does not know.

  4. Step 4: Pick the next test

    We rank the test that would teach the model most.

A test result is one airflow step of an experiment.

Nearby means within one step in oxygen, pressure and airflow combined. All numbers come from the analysis outputs.

Can you trust it?

Scored on ISS tests the model did not see. Details in section 4 below.

How we used SI

SI is the Space Apps word for AI tools. The code, text and analysis were written with SI help (Claude, by Anthropic). Every number comes from NASA data that this code processed; see Sources. The app itself has no generative AI. Its model is a Gaussian-process classifier (a statistical model) trained on NASA data.

Technical details

For reviewers. You do not need these to use the site.

Where the data comes from, what we kept, and what is missing.

How ISS test results become a chance that a flame keeps burning.

How the app picks which hypothetical test to suggest next.

We hid whole ISS experiments, then checked the model's guesses against them. This is not new experiments.

Rerun it and get the same numbers.

Every source is linked. Context is never used to train the model.

Read these before relying on any number.

What SI helped with, and what the app does.