Training an AI Visual Inspector: Citizen Science and Deep Learning for Roman Slitless Spectroscopy

PI: Scarlata, Claudia, University Of Minnesota
Wide-Field Science – Regular

The Nancy Grace Roman Space Telescope is set to revolutionize our understanding of dark energy and cosmic acceleration by measuring weak gravitational lensing and galaxy clustering. A critical requirement for this mission is the accurate redshift determination of millions of galaxies, which will be achieved through slitless spectroscopy. However, this technique, while offering unparalleled multiplexing capabilities by capturing spectra across the entire field of view, faces significant challenges. Overlapping spectra, contamination from adjacent sources, and morphological complexities of galaxies can severely complicate redshift measurements. Traditional 1D extraction methods fall short in addressing these issues, and while forward modeling can improve results, it is computationally prohibitive given the enormous data volumes expected. The work proposed here is designed to tackle these challenges by developing an automated quality validation pipeline that combines the strengths of citizen science with advanced deep learning techniques. Leveraging the Zooniverse platform, the project will engage millions of volunteers to visually inspect and annotate 2D slitless spectroscopic images. These annotations--identifying artifacts, contamination, and real emission line features--will form the training dataset for the project’s deep learning models.

The proposal has three primary objectives:

• Develop an automated pipeline (AISpector) to validate the quality of spectroscopic data by detecting and flagging artifacts.

• Create a non-parametric tool (ELMtool) for continuum and contamination subtraction from 2D spectra, optimized for extracting emission line maps.

• Deploy these tools on the Nexus Roman Science Platform for use by the broader Roman science community.

The plan of execution is structured into three key milestones. Milestone #1 involves launching a citizen science project with two workflows: one for annotating global, detector-level artifacts and another for object-based emission line verification and annotation. Milestone #2 focuses on training two deep learning models using the annotated data. The first model, AISpector (DL#1), is a supervised vision model that flags contaminated regions in 2D spectroscopic images and assign quality flag to automatically identified emission lines. The second model, ELMtool (DL#2), performs non-parametric subtraction of continuum and contamination to optimize emission line extraction. Milestone #3 culminates in the public deployment of these tools on the Nexus platform and demonstrates their scientific utility through applications such as extracting emission line maps from strongly lensed galaxies.

By integrating human expertise with machine learning, Roman’s AISpector and ELMtool aim to enhance the reliability and efficiency of slitless spectroscopic analyses, ensuring that the Roman Telescope’s vast and complex datasets yield the highest quality scientific results.