25–27 Mar 2026
Asia/Taipei timezone

Uncertainty-aware Neural Networks for Fuzzy Dark Matter Model Selection from xHI Measurements

25 Mar 2026, 14:08
3m
Astronomy and Astrophysics Poster Talks

Speaker

Bahareh Soleimanpour Salmasi (National Tsing Hua University)

Description

Fuzzy dark matter (FDM) provides an alternative to the standard cold dark matter picture by suppressing small-scale structure formation through quantum pressure effects. In this work, we test whether FDM can better reproduce the high neutral hydrogen fractions inferred from recent JWST observations at redshifts z > 8. We generate 21 cm reionization histories with 21cmFirstCLASS over a grid of FDM masses and fractions, and incorporate JWST $x_{HI}$ constraints using Bayesian inference that preserves their full non-Gaussian uncertainty distributions. To compare simulations with observations, we use a hybrid neural-network framework that combines convolutional layers for spatial features with recurrent layers for redshift evolution. We find that models near $m_{FDM}$ ~ $10^{-22}$ eV and $f_{FDM}$ ~ 0.04 provide the best agreement with the current data, while lighter masses are more strongly constrained. These results suggest that FDM can delay early structure formation and produce a later, more gradual reionization history, in better agreement with current high-redshift observations.

Author

Bahareh Soleimanpour Salmasi (National Tsing Hua University)

Co-author

Ms S. Mobina Hosseini (Shahid Beheshti University, Tehran, Iran)

Presentation materials