Ultracool dwarf Science with MachIne LEarning (USMILE). I. Scalable Tree-Based Models for Photometric Spectral Classification and New Discoveries from LSST Data Preview 1 and Euclid Quick Data Release 1

TL;DR

The paper tackles the challenge of discovering ultracool dwarfs in the data-rich LSST and Euclid era by building USMILE Avocado, a scalable two-model framework (classifier and regressor) based on gradient-boosted trees that can handle missing photometric data. It leverages a large, uncertainty-augmented labeled set derived from UltracoolSheet, reddened stars, and quasars, with eight colors anchored on $y_{ m LSST}$ computed from synthesized LSST, VHS, and CatWISE photometry, and validated against Euclid QDR1 spectra. The results show ROC AUC $=0.976$, F1 $=0.92$ for classification, and regressor MSE $=0.88$ subtypes, with external spectroscopic validation from Euclid confirming 15 new M6–L2 ultracool dwarfs and guiding reliability regimes, plus 25 additional candidates. The work demonstrates scalable, bias-resistant photometric spectral classification using ML in the LSST-Euclid era, enabling a more complete ultracool-dwarf census and informing follow-up strategies.

Abstract

We present the Ultracool dwarf Science with MachIne LEarning (USMILE), a program developing machine-learning tools for the discovery and characterization of ultracool dwarfs. We introduce USMILE Avocado, a spectral classification framework that uses broadband photometry from wide-field surveys -- Rubin Observatory LSST Data Preview 1, VISTA Hemisphere Survey, and CatWISE -- as input features. The framework has two gradient-boosted decision-tree models scalable to the massive data volumes of modern surveys: the classifier, which distinguishes ultracool dwarfs from stellar/extragalactic contaminants, and the regressor, which predicts spectral types. A key strength is its ability to natively handle missing photometric features, whereas earlier machine-learning approaches required complete multi-band detections or relied on imputation, thereby excluding genuine ultracool dwarfs or introducing bias. Trained on an augmented labeled dataset of >2 million sources built from known ultracool dwarfs, reddened early-type stars, and quasars, the models achieve strong performance: the classifier attains an ROC AUC of 0.976 and an F1 score of 0.92, while the regressor yields a mean-squared error of 0.88 subtypes. Applying these models, we carried out the first ultracool dwarf search with LSST DP1, cross-matched against VHS and CatWISE. Crucially, Euclid Quick Data Release 1 provided near-IR spectra for hundreds of candidates, enabling a rare, large-scale external spectroscopic validation. This confirmed 15 M6--L2 discoveries, verified USMILE performance, and clarified regimes where USMILE predictions are most reliable. Building on these insights, we identified 25 additional M6--L9 photometric candidates. These demonstrate the effectiveness of machine-learning methods in the data-rich era of wide-field surveys, highlighting the synergy between LSST and Euclid in expanding the ultracool dwarf census.

Figures (3)

• Figure 1: Photometric differences between Pan-STARRS1 and LSST in the $i$ (top), $z$ (middle), and $y$ (bottom) bands, derived from synthetic model spectra of the SPHINX (green), Sonora Diamondback (orange), and Exo-REM (blue) grids. Each circle marks an individual synthetic spectrum from the respective model grid.
• Figure 2: The t-SNE projection of the complete labeled dataset (Section \ref{['subsec:classifier']}). Positive (ultracool dwarfs) and negative (reddened early-type stars and quasars) samples are shown in blue and grey, respectively.
• Figure 3: ROC curves for the USMILE baseline classifier (left) and customized classifier (right). The corresponding area-under-the-curve (AUC) values are summarized in Table \ref{['tab:performance']}. Open circles mark the points corresponding to the selected probability threshold for binary classification. In the left panel, the false-positive rate on the x-axis is shown on a logarithmic scale.