Skip to content

Mapping Dark-Matter Clusters via Physics-Guided Diffusion Models

Conference: ECCV 2026
Paper: ECCV 2026 official page (original PDF via virtual poster 5388)
Code: https://graphics.unizar.es/projects/DarkMatterMapping/ (project page, with the method and the DARKCLUSTERS-15K dataset)
Area: Diffusion Model / Scientific Imaging Inverse Problems
Keywords: gravitational lensing, diffusion prior, posterior sampling, dark-matter mapping, cosmological simulation

TL;DR

The paper releases DARKCLUSTERS-15K, the largest galaxy-cluster mass-mapping dataset to date (15,000 paired surface-mass-density maps and simulated three-band HST photometry), trains a photometry-conditioned diffusion prior on it, and turns strong- and weak-lensing physics into likelihood guidance terms driven by DAPS annealing posterior sampling — producing 512×512 non-parametric mass maps with calibrated per-pixel uncertainties in minutes, beating Kaiser-Squires, Napier et al. and a UNet baseline on the TNG/SIMBA held-out test sets and approaching expert-tuned reconstructions on the real cluster MACS 1206.

Background & Motivation

Galaxy clusters are among the most valuable probes in astrophysics and cosmology: their abundance and mass profiles constrain cosmological models directly, and roughly 85% of their mass is dark matter. The only way to actually "see" that dark matter is gravitational lensing — light from background objects is deflected by the curved space-time of the cluster, with the deflection set by the surface mass density along the line of sight. Clusters split into two regimes with very different observational situations. The rare giants at 1e14–1e15 M⊙ offer abundant constraints from both strong lensing (multiple images of the same background source in the dense core) and weak lensing (subtle distortions of background galaxy shapes in the outskirts), and consequently almost all mass-mapping work targets them. The 1e13.5–1e14 M⊙ range is ten to a hundred times more common, yet multiply imaged sources are scarce and weak-lensing shear carries a signal-to-noise ratio of only ~0.01–0.1, so the inverse problem becomes severely ill-posed.

That is precisely the dilemma of current methods. Tools such as Lenstool and WSLAP+ (the latter a parametric plus free-form hybrid) achieve state-of-the-art results, but an expert must decide the number and type of parametric models for every individual cluster (~10–100 parameters for Lenstool, ~1000 for WSLAP+), and end-to-end runtime is measured in hours; stacking many clusters recovers only ensemble-average properties and throws away per-cluster information. Fully automatic alternatives each have their own flaw. Napier et al. extract elliptical light sources from a photometric image with SE++, attach a PIEMD parametric model to each cluster member, and enforce a 1:1 Light-Traces-Mass (LTM) ratio. It finishes in minutes, but hard-wires the mass-to-light relation: once the truth deviates from 1:1, the reconstruction overestimates mass (spurious features) or underestimates it (missing outer mass), and it uses no lensing observables at all. Purely machine-learning routes (a UNet regressing mass from photometry) likewise ignore lensing constraints, cannot constrain the non-luminous dark matter, and output no uncertainty. Upcoming wide-field surveys will discover clusters in this mass range in bulk (Euclid alone found 26 million galaxies in a single week), so per-cluster expert tuning is not sustainable.

Core idea: replace the hard-wired 1:1 parametric mass-to-light assumption with a photometry-conditioned diffusion prior learned from large-scale simulations, write strong and weak lensing as physical forward models with explicit analytic form and inject them as likelihood guidance into the reverse diffusion process, and couple the two through decoupled annealing posterior sampling — yielding non-parametric, physically consistent mass maps together with calibrated per-pixel uncertainties.

Method

The paper solves an inverse problem: given a cluster's photometric images \(P_f\) (three HST bands), the angular positions \(\theta\) of strong-lensing multiple images, and weak-lensing shear measurements \(\gamma\), estimate its dimensionless surface mass density \(\kappa\) (equivalently the surface mass density \(\Sigma\), with \(\kappa=\Sigma/\Sigma_{\rm cr}\), where \(\Sigma_{\rm cr}\) is the critical surface mass density set by the lens and source distances). The difficulty is twofold: lensing observables alone do not uniquely determine the mass distribution, and there is no analytic forward model linking photometry to mass.

Overall Architecture

Three branches feed a single inference procedure. The first is training data: 600 clusters are extracted as 3D point clouds from the TNG300-1 subset of IllustrisTNG and from SIMBA, projected along 25 different lines of sight, yielding paired surface mass density maps \(\Sigma(\theta)\) and three-band photometry \(P_f(\theta)\) that form DARKCLUSTERS-15K (12,000 train / 3,000 test). The second is the generative prior: a DDPM++ score network is trained to model the distribution of these mass maps, but it additionally receives the photometric images as a condition, so what it learns is the joint distribution of light and mass rather than an unconditional mass distribution. The third is the observations and physical constraints at inference time: simulated strong-lensing multiple-image positions, weak-lensing shear (30 sources per square arcminute, with Gaussian noise) and the real photometry \(P_f\) are turned into likelihood terms through the strong- and weak-lensing forward models.

At inference, DAPS (Decoupled Annealing Posterior Sampling) interleaves prior and likelihood: each step first draws an estimate of the clean mass map from a conditional posterior combining the diffusion prior with the likelihood gradient, then re-injects Gaussian noise according to an annealing schedule, driving \(t\) down to 0. Repeating this 20 times yields 20 posterior samples; their mean is the final mass map and their standard deviation is the per-pixel uncertainty.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["IllustrisTNG / SIMBA<br/>600 cluster 3D point clouds"] --> B["DARKCLUSTERS-15K<br/>25 line-of-sight projections<br/>paired Σ and 3-band photometry"]
    B --> C["Photometry-conditioned diffusion prior<br/>DDPM++ learns joint light-mass"]
    D["Observations: photometry P_f<br/>SL multiple images + WL shear"] --> E["Physical forward likelihood terms<br/>SL geometry + WL shear"]
    C --> F["DAPS annealing posterior sampling<br/>prior and likelihood alternating"]
    E --> F
    F --> G["20 posterior samples<br/>mean mass map + per-pixel uncertainty"]

Key Designs

1. DARKCLUSTERS-15K: turning cosmological simulations into paired mass-photometry observations

This is the foundation of the paper and the most direct leap over existing benchmarks. Before it, cluster-scale lensing studies either compared against the two hand-crafted clusters of the Frontier Fields project or borrowed the Camels Multifield Dataset, which targets larger volumes at much coarser resolution — there was no dataset able to support large-scale training and fair benchmarking. The authors take 600 3D clusters (540 from TNG300-1, 60 from SIMBA) as point clouds and project each along 25 distinct lines of sight, producing 25 different lens planes — using viewing angle as data augmentation to inflate 600 simulated objects into 15,000 samples, with the mass and photometry maps recomputed for every projection. Each sample holds four 512×512 images: one surface mass density map \(\Sigma(\theta)\) and simulated HST photometry \(P_f(\theta)\) in three bands (F125W, F606W, F814W), offered in two field-of-view settings, 100×100 and 225×225 arcseconds.

The decisive design choice is that training samples and lensing observations are decoupled. Training and validation use 480 TNG clusters (12,000 projections) at a fixed lens-plane redshift \(z_L=0.5\); the strong- and weak-lensing observations are simulated at inference time only and never enter the diffusion training set. For strong lensing, source angular positions are sampled uniformly on the source plane and the lens equation is solved to find configurations where a single background galaxy is lensed into multiple distinct images — 5 to 20 such sources per cluster depending on total mass. For weak lensing, shear is computed at every angular position from its convolution relation with \(\kappa\), measured at a realistic source density of 30 galaxies per square arcminute, and perturbed with Gaussian noise of variance \(\sigma_w^2\in\{0.03,0.3\}\) (corresponding to shear signal-to-noise ratios of roughly 0.1 and 0.01 respectively), with source redshifts drawn from \(p(z_S)\propto e^{-z_S/z_0}z_S^2/(2z_0^3)\), \(z_0=2/3\), median redshift 1.75. The point of this decoupling is that the model must learn the generalizable task of inferring mass from a class of observations, rather than memorizing what the observation of a given training sample looks like. The test set is deliberately built as two generalization stress tests: 20 TNG clusters rendered at three redshifts \(z_L\in\{0.2,0.5,1.0\}\) with 500 projections each (\(z_L=0.2\) clusters are older with more complex merged structure, \(z_L=1.0\) ones younger and less evolved), and 60 SIMBA clusters (1,500 projections) that use entirely different algorithms — notably for black-hole growth and feedback — to test transfer across simulation frameworks.

2. Photometry-conditioned diffusion prior: replacing the 1:1 LTM assumption with a learned joint light-mass distribution

Parametric methods handle the "light is related to mass but the relation is unknown" problem by simply assuming \(\Sigma\) is proportional to luminosity (LTM, 1:1). The authors consider this too rigid and substitute a diffusion prior: a DDPM++ score network is trained on DARKCLUSTERS-15K, but the key modification is that the score network is additionally conditioned on the corresponding three-channel photometry, so it models not \(\nabla_{x_t}\log p(x_t)\) but the conditional score $$ \nabla_{x_t}\log p(x_t \mid P_f). $$

Concretely the network takes a four-channel image (one mass channel plus three photometry channels), all normalized to \([0,1]\), and is trained with the EDM loss of Karras et al. The motivation rests on a structural observation: photometry has no analytic forward model to mass, so it cannot be placed in the likelihood \(p(y\mid x_0)\) the way lensing can, yet it plainly carries a great deal of information about the dark-matter distribution — which is exactly why the LTM assumption exists. Putting it in the prior's conditioning lets the model learn an arbitrary statistical dependence between light and mass, including a mass-to-light ratio that is not strictly 1:1 and even spatial misalignment between the light and mass distributions, without anyone specifying a functional form. The paper describes this explicitly as learning the statistical relationship between mass and light — a direct replacement for rigid mass-to-light assumptions.

3. Physical forward likelihood terms: wiring strong and weak lensing into the generative process

Lensing observables have a definite physical relationship to mass, so they travel a different route: the likelihood. Rather than modelling the exact likelihood density, the authors drive sampling with its gradient and write the term as a weighted sum of \(k\) loss terms derived from physical forward operators: $$ -\log p(\mathbf{y}\mid\mathbf{x}_0)\ \propto\ \sum_k \lambda_k\,\mathcal{L}_k. $$

For strong lensing, the core requirement is that the lens equation must hold. Since the true angular position \(\beta\) of a background source is unknown, the current mass estimate is used to compute the deflection angle \(\hat\alpha\), a source position is back-solved from each individual image of that source, and the average of those back-solved positions is taken as the source position: $$ \hat{\boldsymbol\beta}{i}=\frac{1}{m_i}\sum}^{m_i}\hat{\boldsymbol\beta{i,j},\qquad \hat{\boldsymbol\beta}), $$ and the geometric loss is the mean squared error between this average and the per-image back-solved positions (averaged over }=\boldsymbol\theta_{i,j}-\hat{\boldsymbol\alpha}(\boldsymbol\theta_{i,j\(n\) sources and each source's \(m_i\) images). There is an easily missed trap here: with the geometric loss alone, all images of all sources collapse onto a single point, since pulling every \(\hat\beta_{i,j}\) together is the cheapest way to satisfy it. The remedy is an additional photometric loss: a parametric elliptical Sérsic profile reconstructs the source at the estimated position \(\hat\beta_i\), and the reconstructed light \(\hat P_f\) is compared against the observed photometry \(P_f\) image by image; this term is small only when the source position and the image positions are both right. In other words, photometry is not treated here as a quantity to be inverted but as a geometric anchor that prevents a degenerate strong-lensing solution.

For weak lensing, two problems arise purely from the observational setup. First, the shear measurements come from sources at different redshifts \(z_S\), and both \(\kappa\) and \(\gamma\) depend on source distance, so a direct pointwise comparison is meaningless; the authors rescale every observation to an arbitrary reference distance \(D_R\) by multiplying and dividing by the appropriate critical density \(\Sigma_{\rm cr}\). Second, shear is measured only at sparse source positions (about 420 measurements are typical for a 225×225 arcsecond field), so a per-pixel data-fidelity term cannot be evaluated; a radial basis function interpolator extrapolates the sparse shear into a dense map covering the whole lens plane, mimicking standard weak-lensing smoothing, which suppresses high-frequency noise while preserving the large-scale signal, and which also lets the likelihood gradient propagate across the entire image rather than only at observed points. Both losses enter the weighted sum above with their own weights. In the \(\kappa\ll1\) regime the authors approximate the reduced shear as \(g=\gamma/(1-\kappa)\approx\gamma\), saving an inversion.

4. DAPS annealing posterior sampling: decoupling prior and likelihood while producing calibrated uncertainties

The conditional diffusion prior and the physical likelihood are combined not by plain classifier guidance but by DAPS. Each update consists of two alternating actions: first, an estimate of the clean mass map \(\tilde{x}_{0|\mathbf y}\) is drawn from a conditional posterior that multiplies the prior term \(p(x_0\mid x_{t+\Delta t})\) supplied by the diffusion model with the physical likelihood term \(p(\mathbf y\mid x_0)\); second, that estimate is perturbed back to a noisy state according to the annealing noise schedule, $$ \mathbf{x}t \sim \mathcal{N}!\left(\tilde{\mathbf x},\ \sigma_t^2 \mathbf I\right), $$ driving \(t\) progressively toward 0. This "draw a clean estimate, then re-anneal with noise" decoupling exists to stop the likelihood gradient, in the high-noise early phase, from dragging the sample into a physically satisfied but prior-implausible corner: the noise-injection step gives the iteration repeated chances to pull the solution back onto the manifold of realistic mass maps via the prior. Because the procedure samples rather than optimizes, one cluster yields many posterior samples whose spread is a per-pixel uncertainty — an essential difference from single-point estimators such as Napier et al. or the UNet, and the reason the authors can report quantities like the total projected mass within 100 kpc of MACS 1206 with an error bar. Relative to parametric methods it is also free-form: 512×512 free parameters, with no NFW- or PIEMD-style shape assumption.

Loss & Training

On the training side there is a single loss: the DDPM++ score network is trained with the EDM recipe on four-channel inputs (mass plus three photometry bands, all normalized to \([0,1]\)). On the inference side, the "loss" is the weighted sum \(\sum_k\lambda_k\mathcal{L}_k\) of the three physical terms — strong-lensing geometric loss, Sérsic photometric loss, and weak-lensing shear loss — each with its own weight \(\lambda_k\); the weak-lensing term is further normalized by the observational noise variance \(\sigma_w^2\) (0.3 for traditional weak-lensing observations, 0.03 for kinematic observations that reduce shape noise using velocity information). Full training hyper-parameters and network details are deferred to the supplemental; the main text does not give a complete hyper-parameter table.

Key Experimental Results

Main Results

Four methods are compared: Kaiser-Squires direct inversion (weak lensing only), the authors' reimplementation of Napier et al. (photometry only with 1:1 LTM, comparable runtime), their own UNet baseline (photometry regressed directly to a mass map), and the proposed method. Metrics include PSNR, SSIM and PCC, plus the RMSE of the deflection angle \(\alpha\), convergence \(\kappa\) and shear \(\gamma\) at held-out points — pixels with neither photometry nor lensing signal — which measures how well the model extrapolates through prior and physics alone where no direct observation exists.

Test set (\(z_L=0.5\)) Metric KS Napier et al. UNet Ours
TNG / SIMBA PSNR ↑ 20.9 / 21.1 26.0 / 22.1 31.1 / 29.6 33.3 / 30.9
TNG / SIMBA SSIM ↑ 0.443 / 0.437 0.703 / 0.565 0.894 / 0.855 0.927 / 0.884
TNG / SIMBA PCC ↑ 0.189 / 0.197 0.832 / 0.616 0.924 / 0.863 0.975 / 0.930
TNG / SIMBA Held-out \(\alpha\) RMSE ↓ 3.697 / 3.919 1.475 / 3.295 0.411 / 1.597 0.530 / 1.485
TNG / SIMBA Held-out \(\kappa\) RMSE ↓ 0.0552 / 0.0521 0.0173 / 0.0169 0.0067 / 0.0073 0.0058 / 0.0067
TNG / SIMBA Held-out \(\gamma\) RMSE ↓ 0.0493 / 0.0437 0.0188 / 0.0217 0.0094 / 0.0174 0.0084 / 0.0151

Except for held-out \(\alpha\) RMSE on TNG, where the UNet is slightly better (0.411 vs 0.530), the method leads on every metric; the advantage survives the jump to the entirely unseen SIMBA framework (PSNR 30.9 vs 29.6, PCC 0.930 vs 0.863), indicating that the prior learns transferable statistical structure rather than TNG-specific texture.

Ablation Study

The ablation feeds one input at a time and then adds the lensing constraints (values read from the annotations in Figure 4 of the paper, ⚠️ refer to the original paper). Two zoomed comparisons are reported:

Input Sample PSNR ↑ Note
Photometry \(P_f\) only Row 1 25.36 Light can only infer dark matter indirectly; the mass estimate overshoots
\(P_f\) + SL + WL (SNR 0.01) Row 1 31.24 Adding lensing constraints: +5.88
Photometry \(P_f\) only Row 2 27.04 Same, the dark-matter component cannot be recovered from light
\(P_f\) + SL + WL (SNR 0.01) Row 2 34.37 +7.33

Key Findings

  • Photometry is necessary but not sufficient. Photometry alone already yields a plausible-looking initial estimate, but since dark matter does not emit light it cannot account for all the mass, and the reconstruction systematically overshoots; strong- and weak-lensing constraints correct this, improving PSNR by more than five points (the paper's own wording, consistent with the 5.88 / 7.33 in the table above).
  • Strong and weak lensing play distinct roles. Strong lensing mainly constrains the surface mass density in the cluster core; weak lensing, despite its noise (SNR 0.01–0.1) and sparse sampling (about 420 measurements typically), still supplies useful information about the outer regions. This matches physical intuition: only cores with \(\kappa\ge1\) produce multiple images, while the outskirts can only be probed through shear statistics.
  • The uncertainties are calibrated, not decorative. A reliability diagram following Kuleshov et al. compares expected against observed confidence levels on a held-out test set, and the agreement between the two indicates that the posterior standard deviation indeed tracks the true error — a capability no pure regression baseline (the UNet) offers.
  • Both failure modes of Napier et al. are visually clear: first, SE++ uses one simple parametric model per detected source, ignoring per-pixel intensity differences, so the mass distribution is over-smoothed; second, once the 1:1 mass-to-light ratio is violated, spurious mass peaks appear in some regions (rows 4 and 5 of the paper's figure) or outer mass is underestimated (row 1). The free-form diffusion prior recovers finer structure in those places.
  • Cross-redshift generalization holds but is not lossless. The model is trained only on TNG clusters at \(z_L=0.5\), yet reconstructs sensibly at \(z_L\in\{0.2,0.5,1.0\}\) and on SIMBA; every SIMBA metric is nonetheless below its TNG counterpart (e.g. PSNR 30.9 vs 33.3), so the difference between simulation frameworks remains a measurable domain gap.

A Real Cluster: MACS 1206

As a proof of concept the pipeline is applied to real data: MACS 1206 (\(z_L\approx0.44\), known for an unusually large number of multiply imaged background galaxies), with three-band HST photometry matching the training bands plus the strong-lensing constraints of Caminha et al. The comparison against the Lenstool expert reconstruction is below:

Metric Caminha et al. (Lenstool, expert-tuned) Ours (fully automatic)
Strong-lensing multiple-image RMS scatter (arcsec) ↓ 0.44 0.56
Total projected mass within 100 kpc (\(10^{13}\,M_\odot\)) \(7.25 \pm 0.02\) \(7.61 \pm 0.04\) (20 samples)
Relative mass error w.r.t. that reference — \(5.03\% \pm 0.49\%\)

The overall halo morphology, orientation and critical curve (at \(z_S\approx1.44\)) agree; the differences concentrate in small-scale peaks near individual cluster members, which the reference's diffuse-halo variant does not model since it ignores light sources in the lens plane. As a reference point, independent expert-tuned models of the same cluster routinely differ by about 10%, so a 5% deviation is well within reason. More importantly, this reconstruction requires no expert tuning and completes in minutes.

Highlights & Insights

  • Put a modality with no analytic forward model into the prior's conditioning instead of inventing a likelihood for it. There is no writable physical equation linking photometry to mass, so rather than forcing a differentiable renderer the authors condition the score network on photometry to learn the joint distribution. This split — "analytic forward model → likelihood; purely statistical association → prior conditioning" — is the methodological backbone of the paper and transfers directly to other scientific inverse problems.
  • Use a photometric loss as an anti-collapse anchor for the strong-lensing geometric loss. The geometric loss alone admits a degenerate solution that pulls every image to one point, a specific and easily missed trap in lensing inversion; bringing photometry in through a Sérsic source reconstruction constrains it cheaply and effectively. Any inverse problem with a degenerate solution can borrow this idea of adding an observable orthogonal to the main residual.
  • Viewing angle as data augmentation. Projecting 600 simulated clusters along 25 lines of sight gives 15,000 samples — extremely cost-effective when N-body simulations are expensive, at the price of samples that are not fully independent.
  • Engineering of the weak-lensing likelihood: rescaling all sources to a reference distance via the \(\Sigma_{\rm cr}\) ratio and then interpolating sparse shear into a dense map with an RBF makes the data-fidelity term differentiable across the whole image. This is the standard combination for turning real observations — heterogeneous in depth, sparse and noisy — into a differentiable loss.
  • Sampling-based inference gives uncertainty for free. The standard deviation over 20 posterior samples serves both calibration evaluation and directly reported physical quantities with error bars, such as the mass within 100 kpc — something parametric single-point estimators cannot do.

Limitations & Future Work

  • The prior's domain gap is a fundamental constraint. All of the model's knowledge comes from TNG300-1 and SIMBA; if the real universe's mass-to-light relation or dark-matter distribution deviates from those simulations, the prior introduces a bias it cannot introspect. On MACS 1206 the agreement is morphological, not a quantitative per-pixel match.
  • The weak-lensing data-fidelity term smooths away high frequencies. The authors themselves call the RBF-interpolated term a "smoothed data-fidelity term" rather than an exact per-pixel noise model, which is lossy for small-scale structure; replacing the true noise covariance with a scalar \(\sigma_w\) is a further simplification.
  • Strong lensing requires paired image positions as input. The method assumes the multiple-image positions and the grouping of images by source are known; on real data this still needs an upstream method or human input, so the boundary of "fully automatic" lies here — the authors do not claim to automate image identification and matching.
  • Training and evaluation share one observation simulator. Both strong and weak lensing observations are generated by the paper's own simulation code; real survey photometric errors, shear measurement biases, non-uniform source density and selection effects are not covered. The cross-simulation (SIMBA) test covers physical modelling differences only, not observational systematics.
  • A single training redshift. The diffusion prior is trained at \(z_L=0.5\) only and generalization to other redshifts relies on the prior's universality rather than explicit conditioning; conditioning on \(z_L\) or the source redshift distribution, or fine-tuning across redshifts, is a natural next step.
  • On resolution, 512 pixels across a 100-arcsecond field is about 0.2 arcsec/pixel and across 225 arcseconds about 0.44 arcsec/pixel (derived by dividing the field evenly; the paper only claims "sub-arcsecond"), still a compromise for the densest cores.
  • vs Lenstool / WSLAP+ (parametric / hybrid, expert-guided): they use parametric halo models such as NFW and PIEMD together with photometry, X-ray or SZ observables and reach state-of-the-art accuracy, but an expert must choose the number and type of models for each cluster, with runtimes in hours; this paper is fully automatic, non-parametric (512×512 free parameters) and finishes in minutes. The trade-off is that its accuracy remains bounded by the simulation prior, whereas parametric methods can write an expert's physical prior directly into the model.
Method Type Photometry SL WL Automation End-to-end time Free params
Lenstool Parametric ✓ (LTM 1:1) ✓ ✓ Expert-guided Hours ~10–100
WSLAP+ Hybrid ✓ ✓ ✓ Expert-guided Hours ~1000
MARS Free-form ✗ ✓ ✓ Fully automatic Hours 100×100
Napier et al. Parametric ✓ (LTM 1:1) ✗ ✗ Fully automatic Minutes ~10–100
Ours Free-form ✓ ✓ ✓ Fully automatic Minutes 512×512
  • vs MARS: MARS is likewise free-form, but uses strong lensing with a simple maximum-entropy regularizer only — no weak lensing and no photometry — and lacks a public implementation; this paper adds both weak lensing and photometry and replaces the hand-crafted regularizer with a learned prior.
  • vs Napier et al.: they use SE++ source extraction plus PIEMD and 1:1 LTM, with photometry only and no lensing observables, so results are over-smoothed and distorted wherever 1:1 is violated; this paper replaces 1:1 with a learnable joint light-mass distribution and puts lensing into the likelihood.
  • vs the UNet baseline: the UNet regresses a mass map directly from photometry — fast, but it cannot use lensing observations (hence cannot constrain dark matter) and produces no uncertainty; this paper lowers the held-out \(\kappa\) RMSE from 0.0067/0.0073 to 0.0058/0.0067.
  • vs Kaiser-Squires and the weak-lensing deep-learning line: KS direct inversion relies on weak lensing alone and is extremely noisy (PCC only 0.19); improvements based on sparsity regularization, inverse-variance filtering or neural score estimation work on large-scale weak-lensing mass mapping at far coarser resolution than this paper's arcminute-scale target. This work sits deliberately between galaxy-scale strong lensing and degree-scale weak lensing.
  • Methodological takeaway: this is another instance of applying DAPS to a scientific inverse problem within the InverseBench framework, showing that the "plug-and-play diffusion prior + physical forward likelihood" paradigm still holds when observational modalities are heterogeneous — those with analytic forward models go through the likelihood, those with only statistical association go into the prior's conditioning.

Rating

  • Novelty: ⭐⭐⭐⭐ First to combine a data-driven score prior with strong plus weak lensing and photometry at cluster scale; the posterior-sampling machinery itself (DAPS) is applied rather than invented, so the contribution lies more in how the problem is recast and in the dataset.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Cross-simulation and cross-redshift test sets, four baselines, ablations, uncertainty calibration and real-data validation are all present; however the weak-lensing noise model is reduced to a scalar and training hyper-parameters are not given in the main text.
  • Writing Quality: ⭐⭐⭐⭐ Physical forward models and losses are laid out term by term, tightly tying formulas to motivation; figures are dense in the conference PDF and some numbers must be read off plots.
  • Value: ⭐⭐⭐⭐⭐ DARKCLUSTERS-15K is the largest galaxy-cluster mass-mapping benchmark to date, and combined with a fully automatic, minute-scale pipeline that reports calibrated uncertainties it has direct practical value for upcoming wide-field surveys.