Title: Bridging Simulations with Observations: Machine Learning and Domain Adaptation Techniques for Cross-Survey Analysis Speaker: Dr. Daniel López (Instituto de Física de Altas Energías (IFAE), UAB, Barcelona, Spain) Date: 2026-02-25 • 15:00 - 16:00 Location: Module 15, Sala 201

Abstract: Modern cosmology relies on combining multiple surveys (DESI, J-PAS, LSST, WEAVE) and state-of-the-art N-body simulations, but differences in data generation, selection, and instrumentation create "domain shifts" that degrade model performance. In this talk, I’ll present a semi-supervised domain adaptation pipeline that bridges this gap for a key task: classifying J-PAS sources into quasars (high/low-z), galaxies, and stars. Accurate quasar classification is crucial—they're rare yet powerful probes of the early universe, and we need clean samples for follow-up surveys like WEAVE-QSO. The challenge: abundant DESI-to-J-PAS mocks don't perfectly match actual observations, while labelled J-PAS data is scarce. My solution pre-trains on large simulations, then adapts using a small set of cross-matched observations, leveraging both statistical power and realism to improve classification where it matters most. I’ll discuss the technical framework, practical lessons for simulation-to-observation transfer, and argue why domain-aware ML pipelines should become standard practice for cross-survey analysis.

Read more: https://weiguangcui.github.io/DECAF/blog#2026-02-25