Speaker
Description
Gaia DR3 provides low-resolution BP/RP (XP) spectra for 220 million sources, with DR4 expected to expand this to around two billion. At XP resolution, narrow features blend together, so standard measurements such as line indices lose sensitivity.
We present GaiaSuperSpec, a neural network trained on stars observed by both Gaia and the Sloan Digital Sky Survey (SDSS). It reconstructs an SDSS-like spectrum from Gaia data, on which standard measurements can be made. Because reconstruction is ill-posed, a network can predict plausible detail not determined by the input. GaiaSuperSpec therefore provides uncertainty estimates through predicted flux ranges and model disagreement.
As a first test, we measured Mgb, an index of the Mg I b triplet strength that was not used to train the model. Using SDSS as the reference, Mgb measured from XP has a mean absolute error of 3.1 Å, while on the reconstructed spectrum the error falls to 0.53 Å. The predicted flux range around Mgb correlates with the Mgb measurement error. A separate regression model trained to predict Mgb from the same Gaia inputs achieves similar accuracy, showing that the information is already present in Gaia data. The reconstruction changes how that information is represented so the standard measurement can recover it.
The same XP, reconstruction, and regression comparison can be applied to Balmer lines, metal and molecular indices, and radial velocity. A key advantage is that the same reconstruction model can be reused for different quantities: regression models must be retrained or extended for each new quantity, whereas one reconstructed spectrum can support many measurements without any of them being specified during training.
If this approach proves reliable, GaiaSuperSpec could make the information encoded in Gaia DR4 XP spectra more accessible to standard spectroscopic measurements and modelling. If GaiaNIR includes low-resolution near-infrared spectrophotometry, an analogous model could be trained using overlap with higher-resolution near-infrared surveys such as APOGEE.