Speaker
Description
Axion dark matter searches such as ABRACADABRA produce continuous high-
rate time series in which injected signals occupy a single narrow frequency bin
per time frame, a structure that is natural in the spectral domain but opaque in
the raw time domain. Existing TIDMAD denoising approaches either operate
directly on raw time series or require separate model weights per frequency band,
limiting generalizability and deployment practicality. We introduce SUNet, a
Spectral UNet that operates on real and imaginary Short-Time Fourier Transform
(STFT) components, enabling inverse-STFT signal reconstruction while training
a single weight set across the full axion frequency range. SUNet is the second
model on TIDMAD, after WaveNet, to generalize across the full frequency range
without per-band retraining, and the first to do so in the spectral domain. Through
frequency-resolved evaluation, we show that SUNet achieves strong recovery from
20 kHz to 4 MHz while degrading below 20 kHz, consistent with elevated 1/f
noise and STFT resolution limits near DC. This characterization, absent from
prior TIDMAD evaluations, reveals frequency-dependent behavior that aggregate
benchmark scores obscure and that matters for targeted axion mass searches.
| Do you plan to submit a 4-page extended abstract on OpenReview (only for Presentations/Posters)? | Yes |
|---|