Deep Needlet: A CNN based full sky component separation method in Needlet space

Debabrata Adak

REVIEW 3 major objections 5 minor 1 cited by

Deep Needlet: A CNN based full sky component separation method in Needlet space

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A U-Net trained on needlet-filtered Planck maps recovers CMB temperature with lower foreground residuals than NILC and matches the true TT spectrum to $\ell \sim 1100$.

desk verdict A genuinely new combination—U-Net on needlet-filtered maps—with solid simulation evidence, but the real-data step leans on training-simulation fidelity that the paper does not fully establish. read the letter →

arxiv 2501.07469 v2 pith:W7THZXMP submitted 2025-01-13 astro-ph.CO astro-ph.GA

Debabrata Adak This is my paper · ORCID

classification astro-ph.COastro-ph.GA

keywords cosmicmicrowavebackgroundcomponentseparationneedletsconvolutionalneuralnetworkU-NetinternallinearcombinationPlanckPR3angularpowerspectrum

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper develops Deep Needlet, a convolutional U-Net that takes needlet-filtered, multi-frequency full-sky maps and outputs a CMB temperature map. The aim is to outperform NILC by letting a trained network learn non-Gaussian foreground morphology instead of applying per-band linear minimum-variance weights. The central claim is that on simulations the recovered map has lower residual foreground contamination than NILC and that its TT power spectrum agrees with the true CMB spectrum up to multipole $\ell \sim 1100$. On Planck PR3 intensity data, the network's map and power spectrum are consistent with the legacy NILC and SMICA products up to $\ell \sim 900$, with the residual small-scale disagreement attributed to instrument systematics absent from training.

What carries the argument

The central object is a set of spherical needlet filters $h^j_\ell$ that satisfy $\sum_j (h^j_\ell)^2 = 1$, decomposing each frequency map into eleven bands localised in both pixel and harmonic space with band-dependent HEALPix resolutions. A U-Net with skip connections is trained separately for each needlet band to map the needlet coefficients of the input frequency maps to the needlet coefficients of the CMB temperature map at 143 GHz. HEALPix maps in nested ordering are rearranged into rectangular grids of size $4 N_\mathrm{side} \times 3 N_\mathrm{side}$, which lets ordinary 2D convolutions approximate rotationally invariant spherical convolution. The cleaned needlet coefficients from all bands are finally recombined through the inverse needlet transform to form the full-sky CMB map.

What would settle it

A concrete falsifier is to run the same trained network on full Planck FFP10 simulations that include instrument systematics and point-source masking, then check whether the recovered TT power spectrum tracks the injected CMB spectrum to $\ell \sim 900$ on the PL76 footprint; if the residual cross-spectrum with input frequency maps stays elevated beyond $\ell \sim 1100$, the claim that reduced residuals come from learned foreground morphology rather than overfitted training features fails.

Watch

Extended reading notes

Core claim

The paper claims that a U-Net operating on band-filtered needlet coefficients of Planck frequency maps separates the CMB temperature signal more accurately than NILC in simulations, recovering the TT power spectrum to $\ell \sim 1100$ with a residual power spectrum below NILC's across the resolved multipole range. When the trained network is applied to Planck PR3 data, the resulting CMB map is consistent with the NILC and SMICA legacy maps, both visually and in power spectrum up to $\ell \sim 900$. The main noticed cost is a small multiplicative bias in the recovered map, attributed to the pixel-wise loss function and corrected with a quadratic fit to the residual power spectrum.

Load-bearing premise

The real-data result depends on the training simulations being close enough to the true sky that the network's learned foreground morphology transfers, yet the training set uses one dust model family, single realizations of CIB, SZ, and CO, no point sources, and no instrument systematics.

Editorial extensions

If this is right

On simulations with foregrounds similar to the training set, Deep Needlet produces CMB maps with smaller residual foreground contamination than NILC, with mean absolute error 4.25 $\mu$K versus 10.26 $\mu$K over the GAL70 mask.
The recovered TT power spectrum matches the true CMB spectrum up to $\ell \sim 1100$ after a quadratic bias correction; below this scale the network systematically underestimates CMB power.
Including LFI channels from 30 to 70 GHz in training reduces the large-scale bias and lowers the mean absolute error to 2.43 $\mu$K, indicating that broader frequency coverage helps the learned foreground subtraction.
On Planck PR3 data, the Deep Needlet map is consistent with the NILC and SMICA legacy CMB maps, and the power spectra agree to $\ell \sim 900$ with the small-scale difference attributed to systematics not included in training.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper, the same band-by-band needlet-space training could be carried over to CMB polarization maps or to future experiments with different frequency sets, since the architecture only needs per-band retraining and a suitable target map.
The persistent negative correlation between the network residuals and the true CMB suggests that a pixel-wise loss is the origin of the bias; augmenting the loss with power-spectrum or summary-statistic terms, as the paper hints, could remove the need for an external bias correction.
The d7 dust test shows that the network also degrades gracefully outside its training foreground distribution, retaining usable accuracy over roughly 70 percent of the sky even when the Galactic-plane residuals grow; sampling multiple CIB, SZ, and dust realizations in training is a direct and testable route to improved generalization.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Bottom line: this is a genuine new combination—a U-Net fed needlet-filtered Planck-frequency maps, with separate networks per needlet band—and the simulation evidence is decent. The authors show lower MAE than NILC, residual power below NILC out to about ell 1100, and the d7 dust swap gives a real out-of-training test. They are also candid about the map-level bias and the small-scale residual. That is more than most ML component-separation papers do.

The soft spots are concentrated in the transfer to real data and in how the headline power-spectrum agreement is produced. The alpha/beta bias correction is fit to the test-simulation reference spectra, so the "agrees to ell~1100" claim is partly by construction; it validates the correction, not the uncorrected map. On the real-data side, training uses a single WebSky realization each of CIB, tSZ, kSZ, one CO template, no point sources, and no systematics. The only out-of-training morphological test swaps d1 for d7 dust; everything else keeps the training template. That does not demonstrate the network has learned to remove real foregrounds generally. The Section 7 attribution of the ell>900 mismatch to "instrument systematics absent from training" is not established—the FFP10 run shows a disagreement at ell<300 instead, so the real-data discrepancy is not cleanly assigned. And the real-data comparison uses cosmic-variance error bars only.

None of this kills the method. The simulation-side claims are reproducible enough to take seriously: held-out realizations, an unseen dust model, and NILC comparison all support the central claim. The concerns are about scope, not soundness. The citation pattern looks appropriate; prior U-Net and needlet work are credited. The main practical frustrations are that code, weights, and seeds are not released, which blocks independent inspection of the exact trained model.

I would send this to review. The method is new, the simulation work is honest, and the real-data weakness is exactly what referees should push on. A careful referee can get this into decent shape. I would cite it if I were working on ML component separation, and I would bring it to reading group.

Referee Report

3 major / 5 minor

Summary. This paper introduces Deep Needlet, a U-Net-style convolutional network that operates on needlet-filtered Planck-frequency maps to recover the CMB temperature map over the full sky. The network is trained on 1000 simulated realizations built from PySM Galactic foregrounds (d1s1f1a2), WebSky CIB and SZ components, a single CO template, and Planck-like noise, with 100 validation and 100 test realizations. The authors report that, on the test simulations, Deep Needlet has lower map-level MAE and lower residual power than NILC, and that after a fitted quadratic bias correction the recovered TT power spectrum agrees with the true spectrum up to approximately ell = 1100. They also test robustness by replacing the d1 dust model with the unseen d7 model, and by adding LFI channels in a second experiment. The trained network is then applied to Planck PR3 intensity data, and the resulting map and power spectrum are compared with the Planck NILC and SMICA products, with agreement reported up to approximately ell = 900.

Significance. If the simulation-based claims hold, the paper is a useful contribution to the CMB component-separation literature: it demonstrates a full-sky, needlet-space CNN that shows lower residual foreground contamination than NILC on held-out simulations, and it is unusually candid about the known biases and limitations of the method. The paper ships no code, but the architecture and data pipeline are described in enough detail to be reproducible, and the held-out test set, the d7 dust swap, and the direct map-level comparison with NILC are genuine strengths. The novelty relative to earlier U-Net CMB work (e.g., Wang et al. 2022) is mostly the needlet-space multi-resolution input representation and the full-sky Planck application, which is incremental but reasonable. The weaker link is the real-data transfer: the training set contains only one realization each of CIB, tSZ, kSZ, and CO, no point sources, and no instrument systematics, so the Section 7 consistency claim rests on an assumption of template transfer that is not directly tested.

major comments (3)

[Section 6, Figure 4] The headline power-spectrum agreement at ell <= 1100 is obtained only after applying the quadratic bias correction <Delta D_ell> = alpha <D_ref> + beta <D_ref>^2, and the text states that the correction parameters are 'estimated in this section' from the same test simulations whose agreement is then quoted. This is a double use of the test set: the corrected spectrum is not an independent prediction, and the reported agreement does not quantify the uncertainty in alpha and beta. Since the same fitted correction is carried over to the d7 test and to the Planck data in Section 7, the real-data spectrum comparison in Figure 10 inherits this issue. Please report uncorrected spectra, cross-validate alpha and beta on an independent simulation set, and propagate their uncertainty into the real-data error bars.
[Section 7, Figure 10] The claim that the Planck PR3 map is consistent with NILC and SMICA up to ell ~ 900 depends on the network generalizing to the true sky, but the out-of-distribution test in Section 6.1 changes only the dust model while keeping the single WebSky CIB/SZ realizations, the single CO template, and the absence of point sources and instrument systematics fixed from training. The Section 7 attribution of the ell > 900 mismatch to 'instrument systematics present in PR3 data' is not established by the FFP10 test described in the same section, which reportedly shows disagreement at ell < 300 rather than at small scales. As written, the real-data agreement could be partly due to foreground-morphology memorization. Please either restrict the real-data claims to the scales where the transfer is explicitly validated, or test the network on a broader set of foreground realizations, multiple CIB/SZ/CO templates, and simulated systematics before claiming consistency with legacy products.
[Section 6, Figure 4, lower-right panel] The central claim that the ML residual power spectrum is lower than the NILC residual 'throughout the scales' is shown only as mean spectra without any uncertainty band or significance estimate. Since the comparison is a quantitative claim and the paper has 100 test realizations available, please add a 1-sigma band or a significance estimate to the residual power spectra so that the reader can judge whether the improvement is statistically meaningful, especially near ell ~ 1000 where the two curves approach each other.

minor comments (5)

[Section 3] The sentence 'We use 143 GHz as the frequency channel for the network’s output' is confusing, because the stated output is the needlet coefficients of the CMB map rather than of a specific frequency channel; please clarify whether this is a typo or whether the network is trained to predict the 143-GHz map as a CMB proxy.
[Section 2, Eq. (2.4)] The text states that the needlet filters satisfy sum_j (h_j^ell)^2 = 1, but the analytic filter shapes in Eq. (2.5) with the band edges in Table 1 do not obviously satisfy this relation without an additional normalization step; please state explicitly how the filters are normalized before use.
[Section 4.2 / Section 7] In Section 7 all power spectra are said to be corrected for their respective beams, but the ML map is already smoothed to a common beam of 7.27 arcmin; please specify whether and how the beam transfer function is divided out of the ML spectrum before the comparison in Figure 10.
[Appendix A, Figure 11] The conclusion that reduced instrument noise does not change the small-scale bias is based on a single noiseless training run; please indicate how many realizations are used in the right panel of Figure 11 and whether the small ell < 200 difference is statistically significant.
[Throughout] There are several typos and label errors, including 'were a_nu is' in Eq. (2.1), 'T able' in the Table 1 caption, and 'foregroud residual' in the Figure 4 legend; a careful proofread would improve the presentation.

Circularity Check

1 steps flagged · score 6.0 of 10

Simulation power-spectrum agreement at ℓ≲1100 is partly in-sample: the two-parameter bias correction is fit to the same test realizations whose spectra are then reported as reconstructed; the map-level NILC comparisons and d7/Planck out-of-sample tests remain independent.

fitted input called prediction [Section 6, paragraphs following Figure 4 (bias-correction discussion); relied on again in Sections 6.1, 6.2, 7 and summarized in Section 8.]
"We correct this bias using a quadratic fit: < ∆Dℓ >= α < DRef erence M ap ℓ > + β < DRef erence M ap ℓ >2 to the residual power spectra as proposed in [50]. This prescription significantly reduced bias at scales ℓ ≲ 1100. ... The same correction approach is applied consistently in Sections 6.1, 6.2, and 7 for simulated and real data respectively, using the correction parameters estimated in this section."

The headline claim that the ML TT power spectrum is accurately reconstructed up to ℓ∼1100 is evaluated on the same 100 test realizations whose residual power spectra were used to fit α and β in ΔDℓ = α D_ref + β D_ref². The corrected spectrum that 'agrees with the theoretical power spectrum' is therefore the residual of an in-sample fit, not an independent prediction of the recovered-map spectrum. The map-level MAE and residual-power comparisons with NILC are held-out and remain independent, and the d7-dust swap and Planck PR3 application use the same fitted correction out-of-sample, so the circularity is partial rather than total. But the specific simulation power-spectrum claim is partly by construction.

full rationale

No self-definitional equivalence, load-bearing self-citation, or uniqueness-imported-from-authors pattern is present: the network is trained on PySM/WebSky simulations and tested on held-out realizations, and the NILC/SMICA comparisons are external benchmarks rather than citations to this paper's own prior work. The only concrete reduction of a headline result to its own inputs is the bias-correction step: α and β are fit to the residual spectra of the very test simulations whose post-correction spectra are then reported as matching the true CMB to ℓ∼1100. That makes the simulation power-spectrum claim partially circular, though the d7-dust test and the Planck PR3 comparison (which reuse the fitted correction on independent data) retain external content. The paper's own disclosure that systematics are absent from training is a limitation on the real-data generalization claim, not an additional circular step; the real-data consistency at ℓ≲900 is an independent comparison even if its error bars (cosmic variance only) are understated. Overall, the central map-recovery demonstration is self-contained and the circularity is localized to the fitted power-spectrum agreement, so the score is 6 rather than higher.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central result rests on standard component-separation and harmonic-analysis assumptions, plus several choices made in this paper: 11 hand-designed needlet bands, a common beam, an architecture with hand-set hyperparameters, and bias-correction coefficients fitted to test simulations. No new physical entities are introduced. The real-data generalization depends on PySM and WebSky simulations being representative, which is only partially tested.

free parameters (5)

Needlet band edges (l_min, l_peak, l_max, Nside) for 11 bands = Table 1; example band 11: 1100, 1500, 1800, Nside 1024
Hand-designed decomposition; defines which angular scales each network branch sees and the pixelization of inputs and outputs.
Quadratic bias-correction coefficients alpha and beta = Not reported numerically
Fitted to the mean residual power spectrum versus reference CMB spectrum in the test set; applied to all reported corrected power spectra including Planck data.
U-Net hyperparameters (kernel size, stride, dropout, batch size, epochs, loss, learning-rate schedule) = Kernel 4x4, stride 2, dropout 0.1, batch 12, 1000 epochs, L1 loss, LR 0.1 to 1e-6
Chosen without an ablation study; they define the trained mapping and the reported performance.
Common beam FWHM = 7.27 arcminutes
Sets the output resolution and effective multipole range; affects all simulated and real maps.
Per-band output normalisation factors = Not specified in the paper
Applied to input and output needlet coefficients to stabilize training; inverse factors are used at testing, but their values are not stated.

assumptions (6)

domain assumption Observed frequency maps are a linear mixture: X_obs = a_nu b CMB + b FG + noise, with CMB uncorrelated with foregrounds and noise.
Standard component-separation model; needed for the network to have a well-defined target. Section 2, Eq. 2.1.
standard math The needlet filters satisfy sum_j (h_j_l)^2 = 1, so forward and inverse needlet transforms preserve power.
Harmonic-analysis partition of unity; used throughout to build inputs and reconstruct the output. Section 2, Eq. 2.4.
domain assumption HEALPix maps arranged as 4 Nside by 3 Nside rectangles give a good approximation to spherical convolution and preserve rotational invariance sufficiently.
Inherited from Wang et al. 2022; Section 5. The paper itself notes band-1 distortion from this approximation.
domain assumption PySM d1s1f1a2 with randomized spectral parameters, plus one WebSky CIB/SZ and one CO realization, spans the foreground diversity of the real sky.
Load-bearing for generalization; tested only against the d7 dust model in Section 6.1, where MAE rises to about 14 microK.
domain assumption Point sources can be excluded from training because the real-data analysis masks them.
Section 4.1 and Section 7; point-source pixels in the output map are inaccurate by construction.
ad hoc to paper The ell greater than 900 disagreement between the real-data ML map and Planck legacy maps is due to instrument systematics absent from training.
Section 7 asserts this might be the cause without a quantitative test; it is an unverified explanation invoked to preserve the central claim on real data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Deep Needlet: A CNN based full sky component separation method in Needlet space." pith.science (2026). https://pith.science/paper/W7THZXMP

@misc{pith2026250107469,
  author       = {Pith},
  title        = {Pith review of: Deep Needlet: A CNN based full sky component separation method in Needlet space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W7THZXMP}},
  note         = {Machine review of arXiv:2501.07469}
}

abstract

One of the key steps in Cosmic Microwave Background (CMB) data analysis is component separation to recover the CMB signal from multi-frequency observations contaminated by foreground emissions. Needlet Internal Linear Combination (NILC) is one of the successful methods that applies the minimum variance estimation technique to a set of needlet-filtered frequency maps to recover CMB. In this work, we develop a deep convolutional neural network (CNN) model to recover CMB temperature map from needlet-filtered frequency maps over the full sky. The network operates on a multi-resolution representation of spherical data, capturing localised features in both pixel and harmonic space, and is designed to preserve the rotational invariance of the CMB signal. The network model is trained on realistic simulations at Planck frequencies, which include CMB temperature maps generated using cosmological parameters sampled within a 2$\sigma$ standard deviation around the Planck best-fit values. We demonstrate the network performance for simulations that exhibit different foreground complexities. The recovered CMB temperature map closely follows the true signal with some residual leakage near the Galactic plane. The TT power spectrum is accurately reconstructed up to multipoles of approximately $\ell\sim 1100$. A minor residual systematics remain at smaller scales. Compared to the NILC method, the network shows reduced residual foreground contamination in the recovered CMB map. Once validated on the simulations, the network is applied to Planck PR3 intensity data. The resulting CMB map is consistent with the CMB maps from the Planck legacy products, including those produced using the NILC and SMICA pipelines. This work demonstrates a powerful component separation method to clean spherical signal data from multi-resolution wavelet-filtered maps.

Figures

Figures reproduced from arXiv: 2501.07469 by the authors.

**Figure 2.** The ML recovered CMB (upper left), NILC recovered CMB (upper right), residual leakage in recovered CMB for ML (lower left) and NILC (lower right). crease in MAE for CMB maps recovered using NILC is only 5%. Additionally, we find a larger bias (< 1.3%) in the power spectrum of the recovered CMB map as compared to the previous configuration. This increment is likely due to the limitation of using multiple realisation… view at source ↗

**Figure 3.** The PDF of the difference between the needlet coefficient maps from network output and reference CMB at needlet bands for one test sample. The differences at bands 1 to 5 are divided by 10 to keep the x-axis range fixed for all bands for easy comparison. 6.1 Testing the impact of dust complexity The network described above was trained using Configuration − 1, which includes the PySM d1 dust model. Since thermal dust… view at source ↗

Figures from the paper (8 more)

**Figure 4.** Figure 4: Upper left panel: Mean TT power spectra from 100 realisations of ML (brown ) and NILC (dark grey) recovered CMB maps and true reference CMB maps (green) estimated over GAL70 mask for foreground configuration comprising only Galactic components. In the bottom panel, we …

**Figure 5.** Figure 5: Left panel: Reference CMB map. Middle panel: ML recovered CMB map using Configuration − 2 as a test simulation. Right panel: Total residual in ML recovered CMB map. – 11 – [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

**Figure 6.** Figure 6: The mean power spectrum of recovered CMB maps using Configuration − 2 as test simulations estimated over GAL70 mask. The green curve is the mean power spectrum of reference CMB maps. The lower panel shows the deviation of the recovered power spectrum from the true powe…

**Figure 7.** Figure 7: Left panel: Bias in recovered CMB maps for experiment − 2 using both cleaning method. The bias in the ML recovered CMB map is considerably reduced for experiment − 2 at ℓ < 1000 as compared to experiment − 1. Right panel: The corresponding residual in the recovered CMB…

**Figure 8.** Figure 8: Same as [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

**Figure 9.** Figure 9: Upper panel: ML recovered CMB from Planck data evaluated using network trained within 30-353 GHz in Section 6.2. The middle and right panel shows the official PR3 CMB map of Planck obtained from NILC and SMICA pipelines, respectively. Lower panel: Pairwise differences …

**Figure 10.** Figure 10: The power spectrum from ML recovered CMB using Planck data between 30-353 GHz estimated over PL76 mask. In the lower panel, we present the deviation from Planck best-fit theoretical power spectrum within a bin of ∆ℓ = 30. The errorbars are cosmic variance at the corre…

**Figure 11.** Figure 11: Left panel: Comparison of bias in ML recovered CMB from trained network with (dark grey) and without (green) instrument noise. Right panel: Corresponding CMB power spectra and residual estimated over GAL70 and corrected for bias. In the left panel of [PITH_FULL_IMAGE…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Single Frequency CMB Foreground Removal with Inter-scale Machine Learning
astro-ph.CO 2026-07 conditional novelty 6.0 of 10

A hybrid CNN using both inter-scale and multi-frequency dust correlations achieves residual B-mode foreground power 3.62e-4 in DustFilaments simulations, about 7x lower than spatial ILC.

Reference graph

Works this paper leans on

73 extracted references · 38 canonical work pages · cited by 1 Pith paper

[1]

T. P. C. P. A. R. Ade, Y. Akiba, A. E. Anthony, K. Arnold, M. Atlas, D. Barron et al., A MEASUREMENT OF THE COSMIC MICROWA VE BACKGROUNDB-MODE POLARIZATION POWER SPECTRUM AT SUB-DEGREE SCALES WITH POLARBEAR , The Astrophysical Journal 794 (2014) 171

work page 2014
[2]

G´ enova-Santos, J

R. G´ enova-Santos, J. A. Rubi˜ no-Mart ´ ın, R. Rebolo, A. Pel´ aez-Santos, C. H. L´ opez-Caraballo, S. Harper et al., QUIJOTE scientific results - I. Measurements of the intensity and polarisation of the anomalous microwave emission in the Perseus molecular complex , MNRAS 452 (2015) 4169 [ 1501.04491]

arXiv 2015
[3]

The Simons Observatory: Project Overview

N. Galitzki, The Simons Observatory: Project Overview , 10, 2018, 1810.02465. – 17 –

work page Pith review arXiv 2018
[4]

Gualtieri, J

R. Gualtieri, J. Filippini, P. Ade, M. Amiri, S. Benton, A. Bergman et al., Spider: Cmb polarimetry from the edge of space , Journal of Low Temperature Physics 193 (2018) 1112

work page 2018
[5]

J. H. Kang, P. A. R. Ade, Z. Ahmed, R. W. Aikin, K. D. Alexander, D. Barkats et al., 2017 upgrade and performance of BICEP3: a 95GHz refracting telescope for degree-scale CMB polarization , in Millimeter, Submillimeter, and Far-Infrared Detectors and Instrumentation for Astronomy IX , J. Zmuidzinas and J.-R. Gao, eds., vol. 10708 of Society of Photo-Optic...

work page Pith review arXiv 2017
[6]

S. K. e. a. Choi, The Atacama Cosmology Telescope: A Measurement of the Cosmic Microwave Background Power Spectra at 98 and 150 GHz , arXiv e-prints (2020) arXiv:2007.07289 [ 2007.07289]

arXiv 2020
[7]

SPTpol Collaborationcollaboration, Measurements of b-mode polarization of the cosmic microwave background from 500 square degrees of sptpol data , Phys. Rev. D 101 (2020) 122003

work page 2020
[8]

Planck Collaboration I, Planck 2018 results. I. Overview and the cosmological legacy of Planck , arXiv e-prints (2018) arXiv:1807.06205 [ 1807.06205]

arXiv 2018

Show all 73 references

[9]

Hanany, M

S. Hanany, M. Alvarez, E. Artis, P. Ashton, J. Aumont, R. Aurlien et al., PICO: Probe of Inflation and Cosmic Origins , arXiv e-prints (2019) arXiv:1902.10541 [ 1902.10541]

2019 arXiv
[10]

Hazumi et al., LiteBIRD: A Satellite for the Studies of B-Mode Polarization and Inflation from Cosmic Background Radiation Detection , J

M. Hazumi et al., LiteBIRD: A Satellite for the Studies of B-Mode Polarization and Inflation from Cosmic Background Radiation Detection , J. Low Temp. Phys. 194 (2019) 443

2019
[11]

D. Adak, A. Sen, S. Basak, J. Delabrouille, T. Ghosh, A. Rotti et al., B-mode forecast of CMB-Bh¯ arat, MNRAS 514 (2022) 3002 [ 2110.12362]

2022 arXiv
[12]

Ichiki, CMB foreground: A concise review , Progress of Theoretical and Experimental Physics 2014 (2014) 06B109

K. Ichiki, CMB foreground: A concise review , Progress of Theoretical and Experimental Physics 2014 (2014) 06B109

2014
[13]

Tegmark, How to measure cmb power spectra without losing information , Phys

M. Tegmark, How to measure cmb power spectra without losing information , Phys. Rev. D 55 (1997) 5895

1997
[14]

Delabrouille, J.-F

J. Delabrouille, J.-F. Cardoso and G. Patanchon, Multidetector multicomponent spectral matching and applications for cosmic microwave background data analysis , Monthly Notices of the Royal Astronomical Society 346 (2003) 1089 [https://academic.oup.com/mnras/article-pdf/346/4/...

2003
[15]

H. K. Eriksen, J. B. Jewell, C. Dickinson, A. J. Banday, K. M. G´ orski and C. R. Lawrence, Joint Bayesian Component Separation and CMB Power Spectrum Estimation , ApJ 676 (2008) 10 [0709.1058]

2008 arXiv
[16]

Delabrouille, J

J. Delabrouille, J. F. Cardoso, M. Le Jeune, M. Betoule, G. Fay and F. Guilloux, A full sky, low foreground, high resolution CMB map from WMAP , A&A 493 (2009) 835 [ 0807.0773]

2009 arXiv
[17]

Basak and J

S. Basak and J. Delabrouille, A needlet internal linear combination analysis of WMAP 7-year data: estimation of CMB temperature map and power spectrum , Monthly Notices of the Royal Astronomical Society 419 (2011) 1163 [https://academic.oup.com/mnras/article-pdf/419/2/1163/311...

2011
[18]

Fern´ andez-Cobos, P

R. Fern´ andez-Cobos, P. Vielva, R. B. Barreiro and E. Mart ´ ınez-Gonz´ alez,Multiresolution internal template cleaning: an application to the Wilkinson Microwave Anisotropy Probe 7-yr polarization data , Monthly Notices of the Royal Astronomical Society 420 (2012) 2162 [http...

2012
[19]

K. K. Rogers, H. V. Peiris, B. Leistedt, J. D. McEwen and A. Pontzen, SILC: a new Planck internal linear combination CMB temperature map using directional wavelets , MNRAS 460 (2016) 3014 [1601.01322]

2016 arXiv
[20]

Remazeilles, A

M. Remazeilles, A. Rotti and J. Chluba, Peeling off foregrounds with the constrained moment ILC method to unveil primordial CMB B-modes, arXiv e-prints (2020) arXiv:2006.08628 [ 2006.08628]

2020 arXiv
[21]

D. Adak, A new approach of estimating the galactic thermal dust and synchrotron polarized emission template in the microwave bands , Monthly Notices of the Royal Astronomical Society 507 (2021) 4618 [https://academic.oup.com/mnras/article-pdf/507/3/4618/40368011/stab2392.pdf]

2021
[22]

Carones and M

A. Carones and M. Remazeilles, Optimization of foreground moment deprojection for semi-blind CMB polarization reconstruction, arXiv e-prints (2024) arXiv:2402.17579 [ 2402.17579]. – 18 –

2024 arXiv
[23]

Stompor, S

R. Stompor, S. Leach, F. Stivoli and C. Baccigalupi, Maximum likelihood algorithm for parametric component separation in cosmic microwave background experiments , Monthly Notices of the Royal Astronomical Society 392 (2008) 216 [https://academic.oup.com/mnras/article-pdf/392/1...

2008
[24]

Basak and J

S. Basak and J. Delabrouille, A needlet ILC analysis of WMAP 9-year polarization data: CMB polarization power spectra, Monthly Notices of the Royal Astronomical Society 435 (2013) 18 [https://academic.oup.com/mnras/article-pdf/435/1/18/3843106/stt1158.pdf]

2013
[25]

J. Kim, P. Naselsky and P. R. Christensen, CMB map derived from the WMAP data through harmonic internal linear combination , PRD 77 (2008) 103002 [ 0803.1394]

2008 arXiv
[26]

J. Dick, M. Remazeilles and J. Delabrouille, Impact of calibration errors on CMB component separation using FastICA and ILC , Monthly Notices of the Royal Astronomical Society 401 (2010) 1602 [https://academic.oup.com/mnras/article-pdf/401/3/1602/3811973/mnras0401-1602.pdf]

2010
[27]

Russell and P

S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach . Prentice Hall, 3 ed., 2010

2010
[28]

H. U. Nørgaard-Nielsen, Foreground removal from WMAP 5 yr temperature maps using an MLP neural network, A&A 520 (2010) A87 [ 1010.1634]

2010 arXiv
[29]

Ravanbakhsh, J

S. Ravanbakhsh, J. Oliva, S. Fromenteau, L. C. Price, S. Ho, J. Schneider et al., Estimating Cosmological Parameters from the Dark Matter Distribution , arXiv e-prints (2017) arXiv:1711.02033 [1711.02033]

2017 arXiv
[30]

Thorne, L

B. Thorne, L. Knox and K. Prabhu, A generative model of galactic dust emission using variational autoencoders, MNRAS 504 (2021) 2603 [ 2101.11181]

2021 arXiv
[31]

Chardin, G

J. Chardin, G. Uhlrich, D. Aubert, N. Deparis, N. Gillet, P. Ocvirk et al., A deep learning model to emulate simulations of cosmic reionization , MNRAS 490 (2019) 1055 [ 1905.06958]

2019 arXiv
[32]

Caldeira, W

J. Caldeira, W. L. K. Wu, B. Nord, C. Avestruz, S. Trivedi and K. T. Story, DeepCMB: Lensing reconstruction of the cosmic microwave background with deep neural networks , Astronomy and Computing 28 (2019) 100307 [ 1810.01483]

2019 arXiv
[33]

Choudhury, A

M. Choudhury, A. Datta and S. Majumdar, Extracting the 21-cm power spectrum and the reionization parameters from mock data sets using artificial neural networks , Monthly Notices of the Royal Astronomical Society 512 (2022) 5010 [https://academic.oup.com/mnras/article-pdf/512/...

2022
[34]

Hassan, S

S. Hassan, S. Andrianomena and C. Doughty, Constraining the astrophysics and cosmology from 21 cm tomography using deep learning with the SKA , MNRAS 494 (2020) 5761 [ 1907.07787]

2020 arXiv
[35]

Nasir, P

F. Nasir, P. Gaikwad, F. B. Davies, J. S. Bolton, E. Puchwein and S. E. I. Bosman, Deep Learning the Intergalactic Medium using Lyman-alpha Forest at 4 ≤ z ≤ 5, arXiv e-prints (2024) arXiv:2404.05794 [2404.05794]

2024 arXiv
[36]

Guzman and J

E. Guzman and J. Meyers, Reconstructing patchy reionization with deep learning , PRD 104 (2021) 043529 [2101.01214]

2021 arXiv
[37]

Chanda and R

P. Chanda and R. Saha, An unbiased estimator of the full-sky CMB angular power spectrum at large scales using neural networks , MNRAS 508 (2021) 4600 [ 2102.04327]

2021 arXiv
[38]

Jeffrey, F

N. Jeffrey, F. Boulanger, B. D. Wandelt, B. Regaldo-Saint Blancard, E. Allys and F. Levrier, Single frequency CMB B-mode inference with realistic foregrounds from a single training image , MNRAS 510 (2022) L1 [ 2111.01138]

2022 arXiv
[39]

McCarthy, J

F. McCarthy, J. C. Hill, W. R. Coulton and D. W. Hogg, Signal-preserving CMB component separation with machine learning , arXiv e-prints (2024) arXiv:2404.03557 [ 2404.03557]

2024 arXiv
[40]

Gawade, A

P. Gawade, A. More, S. More, A. Kimura, A. Sonnenfeld, M. Oguri et al., Neural network prediction of model parameters for strong lensing samples from Hyper Suprime-Cam Survey , arXiv e-prints (2024) arXiv:2404.18897 [2404.18897]

2024 arXiv
[41]

S. Pal, S. K. Yadav, R. Saha and T. Souradeep, Accurate and Unbiased Reconstruction of CMB B Mode using Deep Learning, arXiv e-prints (2024) arXiv:2404.18100 [ 2404.18100]

2024 arXiv
[42]

Yan, G.-J

Y.-P. Yan, G.-J. Wang, S.-Y. Li and J.-Q. Xia, Recovering Cosmic Microwave Background Polarization Signals with Machine Learning , Astrophys. J. 947 (2023) 29 [ 2302.13572]. – 19 –

2023 arXiv
[43]

R. D. Lambaga, V. Sudevan and P. Chen, SkyReconNet: A Deep Learning Inpainting Approach for Enhanced CMB Map Reconstruction, 2501.06139

arXiv
[44]

Marinucci, D

D. Marinucci, D. Pietrobon, A. Balbi, P. Baldi, P. Cabella, G. Kerkyacharian et al., Spherical needlets for cosmic microwave background data analysis , MNRAS 383 (2008) 539 [ 0707.0844]

2008 arXiv
[45]

T. S. Cohen, M. Geiger, J. Koehler and M. Welling, Spherical CNNs, arXiv e-prints (2018) arXiv:1801.10130 [1801.10130]

2018 arXiv
[46]

”Max” Jiang, J

C. ”Max” Jiang, J. Huang, K. Kashinath, Prabhat, P. Marcus and M. Niessner, Spherical CNNs on Unstructured Grids, arXiv e-prints (2019) arXiv:1901.02039 [ 1901.02039]

2019 arXiv
[47]

Krachmalnicoff and M

N. Krachmalnicoff and M. Tomasi, Convolutional neural networks on the HEALPix sphere: a pixel-based algorithm and its application to CMB data analysis , A&A 628 (2019) A129 [ 1902.04083]

2019 arXiv
[48]

K. Yi, J. Chen, Y. G. Wang, B. Zhou, P. Li` o, Y. Fan et al., Approximate Equivariance SO(3) Needlet Convolution, arXiv e-prints (2022) arXiv:2206.10385 [ 2206.10385]

2022 arXiv
[49]

Perraudin, M

N. Perraudin, M. Defferrard, T. Kacprzak and R. Sgier, DeepSphere: Efficient spherical convolutional neural network with HEALPix sampling for cosmological applications , Astronomy and Computing 27 (2019) 130 [ 1810.12186]

2019 arXiv
[50]

M. A. Petroff, G. E. Addison, C. L. Bennett and J. L. Weiland, Full-sky Cosmic Microwave Background Foreground Cleaning Using Machine Learning, ApJ 903 (2020) 104 [ 2004.11507]

2020 arXiv
[51]

Wang, H.-L

G.-J. Wang, H.-L. Shi, Y.-P. Yan, J.-Q. Xia, Y.-Y. Zhao, S.-Y. Li et al., Recovering the CMB Signal with Machine Learning , ApJ 260 (2022) 13 [ 2204.01820]

2022 arXiv
[52]

Ronneberger, P

O. Ronneberger, P. Fischer and T. Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, arXiv e-prints (2015) arXiv:1505.04597 [ 1505.04597]

2015 arXiv
[53]

McCarthy and J

F. McCarthy and J. C. Hill, Component-separated, CIB-cleaned thermal Sunyaev-Zel’dovich maps from Planck PR4 data with a flexible public needlet ILC pipeline , Phys. Rev. D 109 (2024) 023528 [2307.01043]

2024 arXiv
[54]

K. He, X. Zhang, S. Ren and J. Sun, Deep Residual Learning for Image Recognition , arXiv e-prints (2015) arXiv:1512.03385 [ 1512.03385]

2015 arXiv
[55]

D. P. Kingma and J. Ba, Adam: A method for stochastic optimization , CoRR abs/1412.6980 (2014)

2014 arXiv
[56]

Aghanim, Y

Planck Collaboration, N. Aghanim, Y. Akrami, M. Ashdown, J. Aumont, C. Baccigalupi et al., Planck 2018 results. VI. Cosmological parameters , A&A 641 (2020) A6 [ 1807.06209]

2020 arXiv
[57]

Thorne, J

B. Thorne, J. Dunkley, D. Alonso and S. Næss, The Python Sky Model: software for simulating the Galactic microwave sky , MNRAS 469 (2017) 2821 [ 1608.02841]

2017 arXiv
[58]

Planck Collaboration, R. Adam, P. A. R. Ade, N. Aghanim, M. I. R. Alves, M. Arnaud et al., Planck 2015 results. X. Diffuse component separation: Foreground maps , A&A 594 (2016) A10 [ 1502.01588]

2016 arXiv
[59]

Remazeilles, C

M. Remazeilles, C. Dickinson, A. J. Banday, M.-A. Bigot-Sazy and T. Ghosh, An improved source-subtracted and destriped 408-MHz all-sky map , Monthly Notices of the Royal Astronomical Society 451 (2015) 4311 [https://academic.oup.com/mnras/article-pdf/451/4/4311/3903853/stv1274.pdf]

2015
[60]

M. A. Miville-Deschˆ enes, N. Ysard, A. Lavabre, N. Ponthieu, J. F. Mac ´ ıas-P´ erez, J. Aumont et al., Separation of anomalous and synchrotron emissions using WMAP polarization data , A&A 490 (2008) 1093 [0802.3345]

2008 arXiv
[61]

Ali-Ha ¨ ımoud, C

Y. Ali-Ha ¨ ımoud, C. M. Hirata and C. Dickinson,A refined model for spinning dust radiation , MNRAS 395 (2009) 1055 [ 0812.2904]

2009 arXiv
[62]

B. T. Draine and B. S. Hensley, Quantum Suppression of Alignment in Ultrasmall Grains: Microwave Emission from Spinning Dust will be Negligibly Polarized , ApJ 831 (2016) 59 [ 1605.06671]

2016 arXiv
[63]

Abergel, P

Planck Collaboration, A. Abergel, P. A. R. Ade, N. Aghanim, M. I. R. Alves, G. Aniano et al., Planck 2013 results. XI. All-sky model of thermal dust emission , A&A 571 (2014) A11 [ 1312.1300]

2014 arXiv
[64]

Planck Collaboration, P. A. R. Ade, N. Aghanim, M. I. R. Alves, M. Arnaud, M. Ashdown et al., Planck 2015 results. XXV. Diffuse low-frequency Galactic foregrounds , A&A 594 (2016) A25 [ 1506.06660]. – 20 –

2016 arXiv
[65]

J. A. Rubi˜ no-Mart ´ ın, F. Guidi, R. T. G´ enova-Santos, S. E. Harper, D. Herranz, R. J. Hoyland et al., QUIJOTE scientific results - IV. A northern sky survey in intensity and polarization at 10-20 GHz with the multifrequency instrument , MNRAS 519 (2023) 3383 [ 2301.05113]

2023 arXiv
[66]

Hurier, J

G. Hurier, J. F. Mac ´ ıas-P´ erez and S. Hildebrandt,MILCA, a modified internal linear combination algorithm to extract astrophysical emissions from multifrequency sky maps , A&A 558 (2013) A118 [1007.1149]

2013 arXiv
[67]

Stein, M

G. Stein, M. A. Alvarez, J. R. Bond, A. van Engelen and N. Battaglia, The Websky Extragalactic CMB Simulations, JCAP 10 (2020) 012 [ 2001.08787]

2020 arXiv
[68]

K. M. G´ orski, E. Hivon, A. J. Banday, B. D. Wand elt, F. K. Hansen, M. Reinecke et al., HEALPix: A Framework for High-Resolution Discretization and Fast Analysis of Data Distributed on the Sphere , ApJ 622 (2005) 759 [ astro-ph/0409513]

2005 arXiv
[69]

Akrami, M

Planck Collaboration, Y. Akrami, M. Ashdown, J. Aumont, C. Baccigalupi, M. Ballardini et al., Planck 2018 results. IV. Diffuse component separation , A&A 641 (2020) A4 [ 1807.06208]

2020 arXiv
[70]

Planck Collaboration, P. A. R. Ade, N. Aghanim, F. Arg¨ ueso, M. Arnaud, M. Ashdown et al., Planck 2015 results. XXVI. The Second Planck Catalogue of Compact Sources , A&A 594 (2016) A26 [1507.02058]

2016 arXiv
[71]

Tristram, J

M. Tristram, J. F. Mac ´ ıas-P´ erez, C. Renault and D. Santos,XSPECT, estimation of the angular power spectrum by computing cross-power spectra with analytical error bars , MNRAS 358 (2005) 833 [astro-ph/0405575]

2005 arXiv
[72]

III., Planck 2018 results

Planck 2018 results. III., Planck 2018 results. III. High Frequency Instrument data processing and frequency maps, A&A 641 (2020) A3 [ 1807.06207]

2020 arXiv
[73]

T. L. Makinen, A. Heavens, N. Porqueres, T. Charnock, A. Lapel and B. D. Wandelt, Hybrid summary statistics: neural weak lensing inference beyond the power spectrum , arXiv e-prints (2024) arXiv:2407.18909 [2407.18909]. – 21 –

2024 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

T. P. C. P. A. R. Ade, Y. Akiba, A. E. Anthony, K. Arnold, M. Atlas, D. Barron et al., A MEASUREMENT OF THE COSMIC MICROWA VE BACKGROUNDB-MODE POLARIZATION POWER SPECTRUM AT SUB-DEGREE SCALES WITH POLARBEAR , The Astrophysical Journal 794 (2014) 171

work page 2014

[2] [2]

G´ enova-Santos, J

R. G´ enova-Santos, J. A. Rubi˜ no-Mart ´ ın, R. Rebolo, A. Pel´ aez-Santos, C. H. L´ opez-Caraballo, S. Harper et al., QUIJOTE scientific results - I. Measurements of the intensity and polarisation of the anomalous microwave emission in the Perseus molecular complex , MNRAS 452 (2015) 4169 [ 1501.04491]

arXiv 2015

[3] [3]

The Simons Observatory: Project Overview

N. Galitzki, The Simons Observatory: Project Overview , 10, 2018, 1810.02465. – 17 –

work page Pith review arXiv 2018

[4] [4]

Gualtieri, J

R. Gualtieri, J. Filippini, P. Ade, M. Amiri, S. Benton, A. Bergman et al., Spider: Cmb polarimetry from the edge of space , Journal of Low Temperature Physics 193 (2018) 1112

work page 2018

[5] [5]

J. H. Kang, P. A. R. Ade, Z. Ahmed, R. W. Aikin, K. D. Alexander, D. Barkats et al., 2017 upgrade and performance of BICEP3: a 95GHz refracting telescope for degree-scale CMB polarization , in Millimeter, Submillimeter, and Far-Infrared Detectors and Instrumentation for Astronomy IX , J. Zmuidzinas and J.-R. Gao, eds., vol. 10708 of Society of Photo-Optic...

work page Pith review arXiv 2017

[6] [6]

S. K. e. a. Choi, The Atacama Cosmology Telescope: A Measurement of the Cosmic Microwave Background Power Spectra at 98 and 150 GHz , arXiv e-prints (2020) arXiv:2007.07289 [ 2007.07289]

arXiv 2020

[7] [7]

SPTpol Collaborationcollaboration, Measurements of b-mode polarization of the cosmic microwave background from 500 square degrees of sptpol data , Phys. Rev. D 101 (2020) 122003

work page 2020

[8] [8]

Planck Collaboration I, Planck 2018 results. I. Overview and the cosmological legacy of Planck , arXiv e-prints (2018) arXiv:1807.06205 [ 1807.06205]

arXiv 2018