Toward a Spectral Foundation Model

Stellar spectra are the primary way we learn about stars, including their composition, age, and special features like star spots or hidden companion stars. Although we can now observe millions of star spectra, our current methods aren't robust enough to fully explain what this starlight reveals about stars. This project aims to resolve this problem: how to connect the latest computer models of stars with the vast number of stellar spectra we've collected.

The researchers will use new artificial intelligence (AI) methods to create AI models. They will then combine these models with detailed computer simulations to analyze star spectra more effectively. This research aligns with national priorities, using AI to accelerate research and training new experts who are knowledgeable in both scientific research and AI.

The investigators will develop spectral foundational models—robust AI models pre-trained on vast numbers of simplified stellar spectral models. These models will be designed to capture the underlying atomic and plasma physics of stellar atmospheres through carefully crafted training tasks. While simplified one-dimensional physical models assuming spherical symmetry are easy to generate, they lack nuance.

Conversely, detailed three-dimensional models with full non-hydrostatics, though accurate, are resource-intensive and can only be generated sparingly. To address this challenge, the researchers will harness the latest Transformer-based neural network models, establishing methods for pre-training these models with robust unsupervised tasks aimed at teaching the models to grapple with the underlying complex physics in stellar atmospheres.

Building upon the AI foundational model, the team will integrate it with vertex frameworks and probability density estimators with neural networks. The research team will examine two specific science cases. First, they will elucidate star-spot distributions using spectra from the Sloan Digital Sky Survey V (SDSS-V).

Second, they will infer characteristics of stellar populations, including age, mass, and metallicity distributions, from galaxy spectra collected by the Dark Energy Spectroscopic Instrument (DESI).

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.