Speaker
Description
In our previous work, we estimated the masses of galaxy clusters from the PSZ2 full-sky sample using deep learning methods (arXiv:2209.10333). Our approach involved generating a simulated training dataset that closely replicated PSZ2 observations, accounting for instrumental effects such as noise and the point spread function (PSF). However, certain factors, such as point sources, were not included in the simulations, leading to increased scatter in our predictions.
In this new work, we adopt a more robust approach by constructing a training dataset consisting exclusively of clean maps—free from noise and point sources. By leveraging Domain Adaptation and transfer learning techniques, we demonstrate that a model trained on these clean maps can be effectively applied to real observations from the NIKA2 LPSZ sample. This enables us to infer galaxy cluster masses within a simulation-based inference framework.
In this talk, we will briefly discuss:
- The Three Hundred simulation – High-resolution zoom-in simulations of galaxy clusters, essential for generating large datasets of mock observations and training deep learning models.
- Domain shift between simulated and real data – Differences in pixel distributions arise due to uncertainties in baryonic physics and instrumental effects, making direct application of models trained with simulated data to real data challenging.
- Domain adaptation in deep learning – Methods that bridge the gap between simulations and observations by learning invariant latent representations despite domain shifts. We will present successful approaches for implementing these techniques
| Would you be interested in presenting a poster if the conference is oversubcribed? | No |
|---|
