Low-S/N Lyman-alpha cubes now yield gas column-density maps: CDE traces filaments beyond 40 arcseconds in KCWI protocluster data
Synopsis
A team including Caltech researchers trained a two-dimensional convolutional neural network, CDE, to invert low signal-to-noise Lyman-alpha integral-field cubes into line-of-sight integrated gas column-density maps. Training data come from IllustrisTNG100, post-processed with simplified Lyman-alpha radiative transfer and an approximate treatment of KCWI noise. On the simulated test set, CDE reaches a median linear RMSE of 1.48×1020 cm^-2 and a median image SSIM of 0.635, both better than a pixel-wise power-law empirical baseline; applied to KCWI observations of the CC2.2 and Hyperion protoclusters, the predicted maps show filamentary structures extending about 40 arcseconds or more, though the results remain conditional on the adopted simulation and radiative-transfer assumptions.
Interpretation
Lyman-alpha emission is now used to infer diffuse intergalactic-medium gas column-density maps directly, not only surface-brightness maps. Knowledge of large-scale cosmic-web structure had rested mainly on Lyman-alpha absorption tomography, which interpolates one-dimensional absorption spectra along sparse sightlines toward bright background sources and degrades rapidly at high redshift as such sources become rare. Across 1200 simulated test subcubes, CDE gives a median linear RMSE of 1.48×1020 cm^-2, a median RMSE in dex of 0.206 and a median image SSIM of 0.635; on the same test set the pixel-wise power-law empirical model gives 2.29×1020 cm^-2, 3.71 and 0.090.
The network feeds 20 wavelength slices as channels into a two-dimensional convolution instead of convolving in three dimensions. Three-dimensional convolution pays for an extra dimension; in controlled tests with nearly identical architectures, the 2D version cut training time by roughly an order of magnitude while yielding comparable reconstruction. The trade-off rests on the data's extreme resolution anisotropy: at z≈2.3 the line-of-sight scale inferred from the line-spread function is about two orders of magnitude larger than the transverse scale set by the spatial point-spread function.
The training set is simulated and deliberately reproduces KCWI noise and pixel correlations. Lyman-alpha signal is extremely faint, so noise easily blends with genuine emission, making noise modeling decisive for any comparison between simulation and observation. Training inputs come from 750 full-field emission cubes produced by running colt radiative transfer on IllustrisTNG100, from which 6000 subcubes are drawn, injected with Poisson and read noise and passed through a simulated coaddition; the test set holds 1200 subcubes and shares no opposite-view cube with training or validation.
Perspective
The demonstrated scope is limited to KCWI observations of the two massive protocluster fields CC2.2 and Hyperion and to simulated data built on IllustrisTNG100 with simplified radiative transfer; the authors explicitly frame the work as a proof-of-concept feasibility study rather than a definitive measurement of the intergalactic medium.
The reconstructions are conditional on the adopted simulation, radiative-transfer configuration and observational noise model; compact dense clumps are systematically underestimated, and diffuse outskirts and the finest structure are hard to recover at low signal-to-noise; training and test subcubes come from the same 25 pointing centers, which may make performance estimates optimistic; the FABLE cross-simulation test covers only a selected sample from one volume and does not isolate the contributions of subgrid physics, mass resolution and environment.