Variations in the Extinction Curve at High Galactic Latitudes as seen by the Dark Energy Spectroscopic Instrument
The dust extinction curve, characterized by the total-to-selective extinction ratio, $R(V)$, varies with grain size and composition. Accurate $R(V)$ values are crucial for dereddening astronomical colors and understanding the interstellar medium. We measure extinction toward stars with Dark Energy Spectroscopic Instrument (DESI) spectroscopy using a machine-learning-based relative of the "pair method," matching reddened and unreddened stars. We use XGBoost to predict stars' intrinsic colors from their spectroscopic parameters ($T_\mathrm{eff}$, $\log g$, $[\mathrm{Fe/H}]$) and subtract them from observed colors to measure reddening in combinations of the $g$, $r$, $z$, $W1$, and $W2$ bands. This approach directly measures reddening, unlike the common thermal-emission-based method of Schlegel et al. (1998; SFD). The reddening is accurate to roughly 0.02 mag in $g-r$ and $r-z$. We develop a forward model to map $g-r$ and $g-W1$ reddenings to $R(V)$ and find significant variation in $R(V)$, with $σ_\mathrm{NMAD}(R(V)) = 0.23$, in dusty regions ($E(g-r) > 0.1$) at high Galactic latitudes observed by DESI. In this footprint, $R(V)$ is bimodal, with modes at 2.9 and 3.4, likely due to large-spatial-scale correlations in the shape of the extinction curve combined with the relatively small number of distinct sightlines with substantial reddening. We find no strong relationship between $R(V)$ and other tracers of the interstellar medium, such as thermal emission, dust temperature, spectral emissivity index $β$, and $Hα$. However, the error in the SFD-predicted extinction correlates well with $R(V)$, implying that a single optical reddening combined with the far-infrared dust brightness can predict variations in the extinction curve.