FUSION: A PINN to Accurately Model Stellar Parameters from Minimal Input Parameters

Overview

Physics is all about approximation. No one formula can account for every factor and predict any event with full accuracy. Over the years, scientists have used different methods to make these predictions, such as mathematical formulas and statistical models. This has been particularly difficult in fields such as astrophysics, where the subjects being studied are mind-blowing distances away from Earth. In recent years, the advancements in machine learning, specifically in neural networks, allow us to computationally model far more complex relationships than before. The FUSION Stellar Model is designed to model, predict, and classify stellar parameters more accurately than traditional methods. It takes inputs of a star’s Effective Temperature, Luminosity, and Radius, and runs its Physics-Informed Neural Network (PINN) algorithm to find twenty-three other parameters of a star such as Mass and Surface Gravity. It was trained on the Gaia mission dataset with over 403 million stellar parameters. The performance of this model varies depending on the star type and parameter it predicts but has an average accuracy of 90.018% and an average speed of 0.008 seconds per round of predictions. While this model is currently limited to stars on the Hertzsprung-Russell diagram (i.e., the model cannot make accurate predictions on other objects in the universe, such as a neutron star), its main use is to better model the distant lights in our universe.

Competition history

• AJAS 2026 Category not listed

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Physics and Astronomy projects
• AJAS Physics and Astronomy projects
• Every AJAS 2026 project