Cosmoglobe: Unified Multi-Wavelength Sky Model

• Cosmoglobe is a unified Bayesian framework that integrates multi-frequency sky data from legacy and current experiments to enhance CMB and foreground analysis.
• It employs an end-to-end methodology with Gibbs sampling to jointly calibrate instruments and infer astrophysical parameters across various wavelengths.
• Data releases demonstrate resolved systematic discrepancies and improved cosmic parameter estimations, setting new standards for open science and reproducibility.

Searching arXiv for recent Cosmoglobe papers to ground the article. Cosmoglobe is a long-term effort to build one statistically coherent model of the sky across radio, microwave, and infrared wavelengths, and to use that model to learn both about the Milky Way and the early Universe. It grew out of the lessons of WMAP and Planck: all telescopes are looking at the same sky, and the best science follows when their data are analyzed jointly in an end-to-end Bayesian framework rather than through separate, loosely connected pipelines. In practice, Cosmoglobe is simultaneously a scientific collaboration and long-term program, a project to re-analyze legacy and current experiments in a unified way, and a software and data-analysis framework built around the Bayesian Gibbs-sampling code Commander3 (Gerakakis et al., 2022, Martins et al., 18 Apr 2026).

1. Origins, scope, and scientific rationale

Cosmoglobe originated as a generalization of BeyondPlanck, which carried out the first full Bayesian end-to-end re-analysis of the Planck Low Frequency Instrument data. BeyondPlanck focused on a single satellite sub-instrument plus a few external maps, whereas Cosmoglobe aims to become a common platform for joint analysis of many experiments, including Planck LFI, WMAP, Planck HFI, Haslam 408 MHz, and future or current experiments such as LiteBIRD, CMB-S4, SPIDER, and COBE-DIRBE (Gerakakis et al., 2022).

The project’s stated scientific targets include precise measurements of large-scale CMB polarization, including the hunt for primordial B-modes; a detailed multi-component model of the astrophysical microwave sky, including CMB, synchrotron, free-free, spinning dust, thermal dust, and compact sources; and joint exploitation of complementary experiments to break degeneracies between instrument calibration and sky signal (Gerakakis et al., 2022). A broader formulation emphasizes cleaner maps of CMB temperature and polarization, improved determination of cosmological parameters such as the optical depth to reionization $\tau$ and the tensor-to-scalar ratio $r$, and a unified description of Galactic and extragalactic foregrounds from radio through infrared (Martins et al., 18 Apr 2026).

A central premise is that all instruments see the same sky, but over limited frequency ranges and with instrument-specific systematics. This motivates a single global model in which the sky and the instruments are solved together at the time-ordered-data level whenever possible. This suggests that Cosmoglobe is intended not merely as a component-separation pipeline, but as a common inferential layer connecting calibration, mapmaking, astrophysical modeling, and cosmological parameter estimation (Watts et al., 2024).

2. Bayesian end-to-end methodology

Cosmoglobe adopts a Bayesian formulation in which the data are modeled as

$d = s(\omega) + n,$

with posterior

$$P(\omega\mid d) = \frac{P(d\mid \omega)P(\omega)}{P(d)} \propto \mathcal{L}(\omega)P(\omega),$$

where the parameter vector $\omega$ includes both sky and instrument parameters (Gerakakis et al., 2022). This is the core methodological distinction from stage-wise pipelines: gains, beams, noise spectra, bandpasses, and sky component amplitudes and spectral parameters are inferred within one joint probabilistic framework rather than treated as separate sequential products (Martins et al., 18 Apr 2026).

For Planck LFI re-analysis, the time-domain data model is explicitly written as

$\begin{split} d_{j,t} & = g_{j,t}P_{tp,j}\left[ B^{\mathrm{symm}}_{pp',j}\sum_{c} M_{cj}(\beta_{p'}, \Delta_{bp}^{j})a^c_{p'} + B^{\mathrm{asymm}}_{pp',j}\left(s^{\mathrm{orb}}_{j,t} + s^{\mathrm{fsl}}_{j,t}\right)\right] \ &\quad + s^{\mathrm{1Hz}}_{j} + n^{\mathrm{corr}}_{j,t} + n^{\mathrm{w}}_{j,t}, \end{split}$

where $g_{j,t}$ is time-dependent gain, $P_{tp,j}$ is the pointing operator, $B^{\mathrm{symm}}_{pp',j}$ and $B^{\mathrm{asymm}}_{pp',j}$ are symmetric and asymmetric beam operators, $r$0 encodes component SEDs and bandpass effects, and the remaining terms describe orbital dipole, far-sidelobe pickup, a 1 Hz systematic template, correlated noise, and white noise (Gerakakis et al., 2022).

Commander3 implements this through Gibbs sampling, iteratively drawing blocks of parameters from conditional posteriors. The paper describing Cosmoglobe’s open framework summarizes this schematically as updates such as

$r$1

together with analogous updates for noise parameters, beam parameters, and systematic templates (Gerakakis et al., 2022). The resulting output is an ensemble of posterior samples rather than a single map, enabling pixel-resolved uncertainties, covariances between components and instruments, and end-to-end uncertainty propagation to derived quantities such as power spectra and cosmological parameters (Watts et al., 2023).

3. Data releases and major analyses

Cosmoglobe Data Release 1 implemented the first joint analysis of WMAP and Planck LFI time-ordered data within a single Bayesian end-to-end framework. It demonstrated computational feasibility, with one complete WMAP+LFI Gibbs sample costing 812 CPU-hr, of which 603 CPU-hrs were spent on WMAP low-level processing (Watts et al., 2023). The DR1 analysis found that WMAP posterior mean temperature maps and the CMB temperature power spectrum were largely consistent with WMAP9, but that polarization maps exhibited significantly lower large-scale residuals, attributed to a better constrained gain and transmission imbalance model (Watts et al., 2023).

A notable DR1 result was an updated CMB dipole amplitude of

$r$2

which is $r$3 higher than the WMAP9 estimate and $r$4 higher than BeyondPlanck, while being in perfect agreement with the HFI-dominated Planck PR4 result (Watts et al., 2023). The same release reported that the longstanding discrepancy between the WMAP K-band and LFI 30 GHz maps was resolved, and that the W-band polarization sky map, excluded from the official WMAP cosmological analysis, appeared visually consistent with the V-band sky map for the first time (Watts et al., 2023).

A second major DR1 result concerned isotropic cosmic birefringence. Using Cosmoglobe DR1 WMAP polarization maps together with BeyondPlanck LFI products, the analysis constrained a global birefringence angle of

$r$5

from LFI+WMAP only, and found

$r$6

when combining synchrotron-dominated channels with Planck HFI, with the main contribution coming from the LFI 70 GHz channel (Eskilt et al., 2023). The significance of these results was modest, but the central values agreed with dust-dominated HFI-based constraints despite very different instrumental and astrophysical systematics, illustrating the intended cross-validation role of Cosmoglobe’s joint framework (Eskilt et al., 2023).

Data Release 2 extended Cosmoglobe into the infrared through a global Bayesian analysis of COBE-DIRBE time-ordered data, jointly with COBE-FIRAS, Gaia, Planck HFI, and WISE (Watts et al., 2024). This delivered new zodiacal-light-subtracted mission-average DIRBE maps spanning 1.25 to 240 $r$7m and constituted the first consistent unification of the infrared and CMB wavelength ranges into one global sky model covering 100 GHz to 1 $r$8m (Watts et al., 2024).

4. Open science, software, and reproducibility

Cosmoglobe places unusual emphasis on full pipeline release as a condition for reproducibility and long-term data utility. The collaboration argues that providing only raw data and final products is not sufficient to guarantee full reproducibility in the future, because low-level ancillary information, build environments, and internal processing logic are otherwise effectively lost (Gerakakis et al., 2022).

The public release described in the open-science paper includes raw data, source code, parameter files, build scripts, dependencies, and documentation, and is presented as the first publicly released end-to-end CMB analysis pipeline that includes low-level data, source code, configs, and docs (Gerakakis et al., 2022). Commander3 and its build system are organized as a standard software repository with a Fortran source tree, CMake infrastructure, documentation, and auxiliary tooling; installation relies on CMake and GNU Make, with automatic download and build of dependencies such as HEALPix, FFTW3, CFITSIO, HDF5, CAMB, and OpenBLAS (Gerakakis et al., 2022).

The project adopts GPLv3 for Commander and Cosmoglobe, motivated in part by compatibility with HEALPix and the desire to preserve openness of derivative work (Gerakakis et al., 2022). Data distribution is handled through helper utilities and documented directory structures, while a precompiled Docker image is provided for smaller-scale use cases (Gerakakis et al., 2022). This infrastructure is explicitly framed not as ancillary convenience, but as part of the scientific design: a full pipeline release increases data longevity by allowing future reprocessing when better methods, complementary datasets, or more computing power become available (Gerakakis et al., 2022).

This emphasis on openness has organizational implications. Cosmoglobe is described as a community hub with both lightweight participation, in which outside users download code and data for independent analyses, and deeper collaboration, in which external groups integrate proprietary or new datasets into the common framework under memoranda of understanding (Gerakakis et al., 2022). A plausible implication is that Cosmoglobe is intended as a shared infrastructure layer for future CMB and multi-frequency sky experiments, rather than as a mission-specific archive.

5. Foreground and infrared modeling beyond the microwave

Cosmoglobe’s expansion beyond microwave CMB analysis is most visible in DR2 and subsequent foreground papers. The global Bayesian DIRBE analysis produced maps with lower zodiacal light residuals, better determined zero-levels, native HEALPix tessellation with 7' pixel size, nearly white noise at pixel scales, and more complete uncertainty characterization through combined MCMC samples and half-mission maps (Watts et al., 2024). It also identified excess radiation between 4.9 and 60 $r$9m that appears static in solar-centric coordinates, a component treated separately from the parametric zodiacal model (Watts et al., 2024).

Using these reprocessed DIRBE maps, Cosmoglobe DR2 derived new CIB monopole constraints. Positive detections were reported in six of the ten DIRBE bands, including

$d = s(\omega) + n,$0

at 2.2 $d = s(\omega) + n,$1m and

$d = s(\omega) + n,$2

at 240 $d = s(\omega) + n,$3m, while upper limits were set in bands affected by solar-centric excess radiation (Watts et al., 2024). These lower values relative to the official DIRBE release were interpreted as consequences of improved zodiacal light and Galactic foreground modeling (Watts et al., 2024).

The zodiacal-light model itself was rederived in DR2 through joint analysis of DIRBE Calibrated Individual Observations, Planck HFI sky maps, and WISE and Gaia compact object catalogs, while retaining the Kelsall et al. parametric form (San et al., 2024). The resulting DR2 maps have zero-levels lower than the K98 zodiacal-light-subtracted mission-average maps by about $d = s(\omega) + n,$4 kJy/sr at 1.25–3.5 $d = s(\omega) + n,$5m, and map RMS values at wavelengths up to 25 $d = s(\omega) + n,$6m that are about $d = s(\omega) + n,$7 lower at high Galactic latitude than the corresponding DIRBE ZSMA maps (San et al., 2024).

A related DR2 paper modeled starlight in DIRBE with Gaia and WISE. It fit 424,829 bright sources individually using Gaia stellar parameters and PHOENIX SEDs, while the remaining 710,825,587 WISE sources were coadded into a diffuse background template with one overall amplitude per DIRBE band (Galloway et al., 12 Jan 2026). The resulting model found that total stellar emission accounts for $d = s(\omega) + n,$8 of the observed flux density at 2.2 $d = s(\omega) + n,$9m, $$P(\omega\mid d) = \frac{P(d\mid \omega)P(\omega)}{P(d)} \propto \mathcal{L}(\omega)P(\omega),$$0 at 4.9 $$P(\omega\mid d) = \frac{P(d\mid \omega)P(\omega)}{P(d)} \propto \mathcal{L}(\omega)P(\omega),$$1m, and $$P(\omega\mid d) = \frac{P(d\mid \omega)P(\omega)}{P(d)} \propto \mathcal{L}(\omega)P(\omega),$$2 at 25 $$P(\omega\mid d) = \frac{P(d\mid \omega)P(\omega)}{P(d)} \propto \mathcal{L}(\omega)P(\omega),$$3m (Galloway et al., 12 Jan 2026).

Thermal dust modeling was likewise reformulated in physically structured terms. A low-resolution analysis showed that large-scale thermal dust emission between 353 GHz and 25 THz can be represented as a linear mixture of five physically motivated ISM tracers with spatially isotropic SEDs, capturing $$P(\omega\mid d) = \frac{P(d\mid \omega)P(\omega)}{P(d)} \propto \mathcal{L}(\omega)P(\omega),$$4 of the full signal RMS below 1 THz (Gjerløw et al., 12 Jan 2026). A higher-resolution Planck HFI analysis then distilled this into a four-component model of cold dust, hot dust, nearby dust, and H$$P(\omega\mid d) = \frac{P(d\mid \omega)P(\omega)}{P(d)} \propto \mathcal{L}(\omega)P(\omega),$$5-correlated dust, capturing over $$P(\omega\mid d) = \frac{P(d\mid \omega)P(\omega)}{P(d)} \propto \mathcal{L}(\omega)P(\omega),$$6 of the full-sky dust variance for all channels while using fewer degrees of freedom per pixel than the Planck 2015 legacy dust model (Sullivan et al., 15 Jan 2026). The corresponding DIRBE analysis showed that the same four components extend into the infrared and that their combined SED agrees well with recent astrodust predictions, supporting the notion of a microwave-to-infrared concordance dust model (Gjerløw et al., 12 Jan 2026).

6. Computational scaling, future directions, and significance

Cosmoglobe’s methodology is computationally intensive by construction, and several papers make scalability itself a scientific topic. DR1 established feasibility for a joint WMAP+LFI analysis at 812 CPU-hr per full Gibbs sample (Watts et al., 2023). A later study assessed the computational feasibility of end-to-end Bayesian analysis of LiteBIRD simulations within Commander3 and estimated that one Gibbs sample from TOD to cosmological parameters would cost approximately 3000 CPU hours (Aurvik et al., 7 Jul 2025). For the full three-year LiteBIRD mission, the paper estimated a total data volume of 238 TB uncompressed, or 70 TB after Huffman compression, and concluded that these requirements are well within capabilities of future high-performance computing systems (Aurvik et al., 7 Jul 2025).

Cosmoglobe also continues to expand its astrophysical scope. A 2026 Commander-based all-sky model of Galactic emission at radio and microwave frequencies incorporated recent surveys such as S-PASS, C-BASS, and QUIJOTE together with reprocessed WMAP and Planck LFI data from Cosmoglobe and Planck HFI channels (Hoerning et al., 19 Jun 2026). The analysis fit 35 full- and partial-sky maps at 1 degree resolution and reported RMS temperature residuals below 10 $$P(\omega\mid d) = \frac{P(d\mid \omega)P(\omega)}{P(d)} \propto \mathcal{L}(\omega)P(\omega),$$7 over 95% of the sky up to 353 GHz, with residual angular power spectra more than two orders of magnitude below the CMB spectrum (Hoerning et al., 19 Jun 2026). Although not itself a formal DR paper, it is explicitly positioned as a Cosmoglobe-style Commander analysis that extends the project’s low-frequency foreground program (Hoerning et al., 19 Jun 2026).

Taken together, these developments indicate that Cosmoglobe is evolving from a reprocessing effort centered on WMAP and Planck into a broader multi-experiment, multi-frequency inference platform. This suggests a dual significance. Scientifically, it seeks cleaner CMB and foreground products with end-to-end uncertainty propagation. Methodologically, it advances a model in which legacy and future datasets remain scientifically active because their raw data, software, and low-level calibration logic are preserved within a common open framework (Gerakakis et al., 2022, Martins et al., 18 Apr 2026).