SPHEREx Ultracool Dwarf Spectral Atlas

• The atlas provides a homogeneous, continuous 0.75–5.0 μm spectral dataset of 1,675 ultracool dwarfs, enhancing comparative population studies.
• It employs uniform spectral fitting and Bayesian refinement to accurately infer atmospheric parameters, bolometric luminosities, masses, radii, and ages.
• The empirical atlas reveals key molecular diagnostics such as CO and CO₂ features and highlights systematic offsets between atmospheric and evolutionary gravities.

The SPHEREx Ultracool Dwarf spectral Atlas (SUDA) is a homogeneous spectroscopic reference sample of ultracool dwarfs constructed from continuous 0.75–5.0 $\mu$m SPHEREx QR2 spectroscopy and designed to connect observed spectral morphology with atmospheric and evolutionary properties across the coolest stellar and substellar regimes (Tu et al., 29 Apr 2026). Its defining feature is the use of survey-wide, continuous SPHEREx spectra rather than heterogeneous stitched spectral energy distributions, which preserves the 3–5 $\mu$m region that contains a large fraction of ultracool-dwarf flux and the CO/CO$_2$ diagnostics often absent from earlier large samples. SUDA combines uniform spectral fitting, bolometric luminosity estimation, evolutionary inference, and an empirical spectral atlas spanning approximately $T_{\mathrm{eff}} \simeq 700$–3000 K.

1. Survey basis and sample construction

SUDA was assembled from UltracoolSheet, a catalog of known ultracool dwarfs and brown dwarfs. Source positions were propagated to the SPHEREx observing epoch using proper motions and then positionally matched to the SPHEREx QR2 archive. This produced spectra for 3164 ultracool objects, of which 1675 remained in the final SUDA analysis sample after quality cuts, astrophysical vetting, and visual and photometric consistency checks (Tu et al., 29 Apr 2026).

The screening was intentionally strict. Only spectral points with the SPHEREx quality flag $2^{21}$ were retained, and spectra with median S/N $<2$ were removed. Unresolved binaries or multiples, problematic sources, and young-region contaminants were excluded. Synthetic broadband photometry derived from the SPHEREx spectra was compared against 2MASS/MKO $JHK$ and WISE $W1$, and obviously contaminated or miscentered sources were rejected. Spectra affected by terrestrial atmospheric emission, especially the He line at 1.083 $\mu$m, were truncated shortward of 1.2 $\mu$m.

Stage	Number of objects
Positional match to SPHEREx QR2	3164
Final SUDA analysis sample	1675

The resulting sample is large and relatively clean while maintaining homogeneous spectral coverage over the full 0.75–5.0 $\mu$0m interval. A plausible implication is that the atlas is better suited to comparative population studies than legacy compilations assembled from heterogeneous instruments and reduction pipelines.

2. SPHEREx spectral coverage and external validation

SPHEREx provides six continuous bands spanning 0.75–5.0 $\mu$1m. In SUDA, the first four bands cover 0.75–3.82 $\mu$2m at $\mu$3, and the two redder bands cover 3.82–5.0 $\mu$4m at $\mu$5–130 (Tu et al., 29 Apr 2026). This instrumental configuration is especially consequential for ultracool dwarfs because a substantial fraction of their emergent flux lies longward of 2.5 $\mu$6m, and key molecular signatures, notably CO and CO$\mu$7, reside in the 3–5 $\mu$8m window.

All SUDA spectra are drawn from SPHEREx QR2, described as the reprocessed archive with improved calibration and data quality relative to QR1. The paper also reports external validation against JWST spectra for three benchmark objects. In those comparisons, agreement in both morphology and absolute flux is described as excellent, with bolometric flux differences of 1.4%, 1.7%, and 4.0%. This establishes that the QR2 spectra are sufficiently stable for integrated-flux work and for model-data comparison in the regime where low-resolution infrared morphology carries much of the atmospheric information.

The emphasis on continuous coverage is methodologically important. SUDA is not organized around isolated diagnostic bands alone; instead, it uses a survey product that samples the full near- to mid-infrared spectral shape in a single homogeneous framework.

3. Atmospheric parameter inference

The atmospheric analysis follows a two-stage procedure: initial $\mu$9 grid fitting and subsequent Bayesian refinement with nested sampling. The primary grid is SAND (Spectral ANalog of Dwarfs), chosen because it spans $_2$0–4000 K together with broad $_2$1 and metallicity coverage. For cooler objects, SUDA also fits ATMO2020++, which is optimized for late-T/Y dwarfs and spans $_2$2 to $_2$3 (Tu et al., 29 Apr 2026).

The model-selection logic is explicit. If a source had a preliminary SAND solution with $_2$4 K, the fit was rerun with ATMO2020++, and the adopted model family was the one with the smaller reduced $_2$5. Before fitting, model spectra were convolved to SPHEREx resolution and scaled by a wavelength-independent factor. The best-fitting grid point from this stage supplied the initial solution.

Final parameter inference used UltraNest nested sampling. The relevant model family was taken from the initial grid fit, and interpolation was performed linearly in all available dimensions in log-flux space. For SAND, the native low-temperature grid reaches $_2$6–6.0; for some $_2$7 K objects initially at $_2$8, the grid was extended by linear extrapolation to $_2$9, but not below that. The likelihood incorporated an effective variance with a fractional error inflation term,

$T_{\mathrm{eff}} \simeq 700$0

and

$T_{\mathrm{eff}} \simeq 700$1

The priors were centered on the initial $T_{\mathrm{eff}} \simeq 700$2 solution: $T_{\mathrm{eff}} \simeq 700$3 K, $T_{\mathrm{eff}} \simeq 700$4 dex, uniform priors over the full grid for $T_{\mathrm{eff}} \simeq 700$5 and $T_{\mathrm{eff}} \simeq 700$6, a prior on $T_{\mathrm{eff}} \simeq 700$7 centered on the initial flux scaling, and $T_{\mathrm{eff}} \simeq 700$8. Posterior medians and 16th/84th percentiles define the reported parameter estimates and uncertainties. This makes the SUDA atmospheric catalog a model-dependent but internally uniform inference product.

4. Bolometric luminosities, radii, masses, and ages

Bolometric fluxes were computed by integrating a composite spectral energy distribution consisting of observed SPHEREx data where available and the best-fit model outside the observed range (Tu et al., 29 Apr 2026). For SAND, the model already spans approximately 0.1–999 $T_{\mathrm{eff}} \simeq 700$9m, so no additional long-wavelength extrapolation is required. For ATMO2020++, the spectra were extrapolated linearly for $2^{21}$0 $2^{21}$1m and with a Rayleigh–Jeans tail for $2^{21}$2 $2^{21}$3m.

For sources with parallaxes, bolometric luminosities were obtained from bolometric flux and distance, with uncertainties propagated through Monte Carlo sampling and including a 5% absolute calibration term. For sources without parallaxes, luminosity could not be obtained directly from flux, so the study trained an XGBoost regressor to predict $2^{21}$4 from $2^{21}$5. The training sample contained 1041 parallax-based objects with reasonable radii of $2^{21}$6–$2^{21}$7, and the final prediction set contained 550 no-parallax objects. The out-of-fold mean absolute error was about 0.12 dex, and the training-set MAE was 0.092 dex.

For parallax-bearing sources, the fitted $2^{21}$8 scaling combined with the distance yielded the radius. These radii, together with $2^{21}$9, were compared to the C23 evolutionary tracks to infer mass, age, and evolutionary $<2$0. The paper states that $<2$1 was deliberately preferred over $<2$2 because it reduces covariance in the evolutionary interpolation. Local affine extrapolation was allowed if a point lay just outside the convex hull of the model grid.

Typical uncertainties were reported as a median mass uncertainty of approximately 21% and age uncertainty of approximately 0.24 dex for the full sample, and a median mass uncertainty of approximately 16% and age uncertainty of approximately 0.15 dex for the parallax subset. These values indicate that SUDA provides a statistically useful evolutionary characterization across a large sample, though the inferred quantities remain conditioned on the adopted atmospheric fits and evolutionary tracks.

5. Atmospheric versus evolutionary surface gravity

One of SUDA’s central empirical results is a discrepancy between atmospheric $<2$3 inferred from spectral fitting and evolutionary $<2$4 inferred from the C23 tracks. The strongest effect occurs at $<2$5–2500 K, where the atmospheric gravities are systematically lower than the evolutionary gravities, with a median offset of about

$<2$6

(Tu et al., 29 Apr 2026).

The paper argues that this is likely not a genuine physical difference. Instead, it attributes the offset to a fitting degeneracy in low-resolution SPHEREx spectra: in this temperature regime, gravity-sensitive changes in L-dwarf spectra are subtle and appear mainly in the $<2$7 band, so they can be partially absorbed by metallicity variations within the broad SAND grid. In the paper’s formulation, $<2$8 and $<2$9 are not cleanly separable in the 1700–2500 K regime with these spectra alone.

This result has methodological significance beyond SUDA itself. It cautions against reading atmospheric $JHK$0 from low-resolution infrared fitting as a direct surrogate for evolutionary gravity without external constraints, particularly in the L-dwarf regime. It also sharpens the distinction between model-fit parameters and physically cross-validated parameters derived from radius and evolutionary interpolation.

6. Empirical atlas and molecular-sequence behavior

A major product of SUDA is an empirical spectral atlas constructed from the observed SPHEREx spectra of objects with parallaxes. Sources were grouped by atmospheric $JHK$1 and evolutionary $JHK$2, using bins of $JHK$3 K and $JHK$4. The spectra were interpolated onto a common wavelength grid, normalized in the 1.27–1.33 $JHK$5m window, and combined via the wavelength-by-wavelength median. Bins containing fewer than five objects were excluded. The final atlas contains 52 parameter bins built from 968 parallax-bearing spectra and spans approximately $JHK$6–3000 K (Tu et al., 29 Apr 2026).

The atlas is explicitly empirical rather than a model-template library. Its dominant trend is temperature-driven: as $JHK$7 decreases, molecular absorption strengthens and broad-band morphology changes systematically from late-M through T/Y types. Gravity dependence is weaker, especially above approximately 1800 K. This suggests that, at SPHEREx resolution and over this wavelength range, temperature dominates the organization of the ultracool-dwarf sequence.

The paper also measures four molecular indices—H$JHK$8O, CH$JHK$9, CO, and CO$W1$0—using the continuum-to-feature ratio

$W1$1

retaining only measurements with $W1$2. H$W1$3O and CH$W1$4 strengthen monotonically toward lower $W1$5, with a particularly steep rise between approximately 1300 and 1100 K, consistent with the rapid chemical transition across the L/T regime as methane becomes favored and water bands deepen. SAND and ATMO2020++ broadly agree for H$W1$6O and CH$W1$7, and both reproduce the overall temperature trend reasonably well.

CO and CO$W1$8 exhibit a different pattern: little overall temperature dependence above approximately 1500 K, a rise below approximately 1500 K, a maximum around approximately 1000 K, and a decline again at lower temperatures. CO$W1$9 shows the larger amplitude and the stronger scatter at fixed temperature. The paper interprets this nonmonotonic behavior as reflecting changing thermal chemical structure through the L/T transition, carbon partitioning into CH$\mu$0 at the cold end, and, in some objects, vertical mixing that maintains stronger CO than equilibrium would predict.

A key empirical conclusion is that the CO$\mu$1 index is especially useful as a metallicity tracer in the approximate range $\mu$2–1300 K. In that interval, the model sequences separate clearly by metallicity, and the observed scatter in CO and CO$\mu$3 is suggested to be substantially metallicity-driven. The paper further states that ATMO2020++ matches the observed CO and CO$\mu$4 trends better than SAND, especially for these carbon-bearing bands.

7. Catalog products, comparative performance, and scientific role

SUDA reports several quantitative comparisons against external or literature benchmarks. For 592 matched objects, the luminosities show good agreement with the literature. Within the 475 parallax-based overlaps, the mean difference is 0.014 dex with MAD 0.020 dex; for the 117 objects without parallax, the machine-learning luminosities differ by a mean of 0.080 dex and MAD 0.090 dex relative to literature values (Tu et al., 29 Apr 2026). Spectral typing based on SPLAT agrees with literature types at roughly the one-subtype level, with a mean offset of $\mu$5 subtype and MAD 0.6.

The catalog table includes fitted $\mu$6, $\mu$7, $\mu$8, $\mu$9, $\mu$0, reduced $\mu$1, best-fit model family, distance, literature and re-derived spectral types, radius, mass, age, evolutionary $\mu$2 where available, $\mu$3, $\mu$4, and a flag for evolutionary-grid extrapolation. The breadth of these outputs makes SUDA simultaneously a spectral archive, a parameter catalog, and a population-level reference set.

Its scientific role follows directly from that structure. SUDA provides a single homogeneous observational bridge linking full 0.75–5 $\mu$5m spectral morphology, uniformly derived atmospheric parameters, bolometric luminosities, evolutionary mass-age-gravity estimates, and molecular chemistry diagnostics. This suggests immediate utility for calibrating ultracool-dwarf atmosphere models, studying the L/T transition and late-T chemistry, identifying metallicity-sensitive observables, and mapping spectroscopic appearance onto the physical parameters that govern atmospheric structure and substellar cooling.