Teaching a neural network to read blurry space spectra directly

Slitless spectrographs smear a galaxy's spectrum across the detector in two dimensions, tangling spectral information with spatial shape and making wavelength calibration messy. Using mock CSST observations built from real HSC imaging and DESI spectra, a Bayesian convolutional network with Monte Carlo dropout reads redshifts directly from those 2D images. At signal-to-noise above 10, the method reaches σ_NMAD = 0.0024 — precise enough for baryon acoustic oscillation science — and spatial augmentation during training makes it robust against calibration offsets. The approach could streamline analysis pipelines for any future wide-field slitless survey.