{"id":"17a159ed-f200-4b25-98fc-b2ca92c9cc21","arxiv_id":"2502.07473","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A multiscale Bayesian estimator using Poisson statistics and simulated X-ray spectra produces less noisy, higher-resolution elemental maps from sparse STEM-EDX data.","lead":"This paper presents a Bayesian method that denoises and quantifies energy-dispersive X-ray maps by combining Poisson statistics, spatial smoothing, and simulated X-ray spectra. It reports sharper atomic-resolution chemical maps of a ferroelectric PbTiO3 sample, including unit-cell-level ferroelectric domain patterns.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The accuracy claim is supported only by an ESPM self-consistency loop; experimental absorption/channelling caveats leave external quantitative accuracy unvalidated.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the ESPM-simulated endmembers are assumed correct, while thin-sample and channelling limitations are acknowledged. My reading of the paper confirms this is the most consequential gap. The method is otherwise well supported: the multiscale idea is plausible, the code is available, the simulated tests are internally consistent, and the experimental comparison for spatial resolution (FFT spots, Atomap column positions, c/a ratio) is a useful demonstration. However, the abstract's accuracy claim is stronger than what the evidence supports, because the only quantitative accuracy test uses the same forward model for data generation and inference. This does not make the method wrong; it makes the accuracy claim conditional on endmember fidelity. The recommended verdict therefore remains CONDITIONAL, and the concrete test above would directly determine whether the concern lands.","tokens_in":14826,"tokens_out":5902,"duration_ms":64403,"concrete_test":"Generate the same simulated PbTiO3/SrRuO3 phantom with an independent forward model (e.g., DTSA-II or a measured reference-standard spectrum) instead of ESPM, and run the RMB algorithm using the ESPM endmembers while keeping all other settings fixed. If the recovered at.% values differ from ground truth by more than the RMB per-pixel uncertainties, then the accuracy claim is not robust to endmember specification and the abstract should be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim rests on simulated data generated with the ESPM library and then quantified with ESPM endmembers. This is an internal-consistency test: it validates that the algorithm can invert the same forward model used to create the data, but it does not validate the physical fidelity of those endmembers. The paper itself limits the model to 'relatively thin samples, where absorption does not represent a relevant contribution' and notes that atomic-resolution EDX concentrations are 'not always directly meaningful, owing to electron channelling'. The experimental PbTiO3 film is 28 nm thick and imaged at atomic resolution, so both effects are plausibly present. If the ESPM cross-sections, Bremsstrahlung background, or detector response are biased for this geometry, the reported abundance maps inherit that bias even though the spatial denoising works. The unit-cell-resolution demonstration is less affected because it uses Pb-column positions rather than absolute concentrations, but it does not establish the unqualified 'more accurate chemical maps' claim. The spatial-resolution gain on experimental data is a genuine positive, but the quantitative accuracy claim needs an external anchor.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a robust multiscale Bayesian (RMB) method for quantitative analysis and denoising of energy-dispersive X-ray (EDX) spectrum images. The method models X-ray counts with a Poisson likelihood, approximates it via a gamma surrogate that uses initial abundance estimates from the SUNSAL unmixing algorithm, and combines this with a multiscale spatial-smoothness prior and an inverse-gamma variance prior. Inference is performed by coordinate-descent MAP estimation. The method is tested on simulated 90x90 spectrum images of PbTiO3/SrRuO3 layers at 5, 50, and 500 counts per pixel, where it is compared with standard Cliff-Lorimer quantification after kernel smoothing, and on an experimental atomic-resolution PbTiO3 spectrum image. In simulations, RMB yields narrower composition histograms, better SNR versus feature broadening, and fewer Otsu-threshold misclassifications. In the experimental data, RMB produces sharper elemental maps, shows additional FFT spots, improves atomic-column localization relative to HAADF positions, and allows unit-cell c/a ratio maps to be extracted.","tokens_in":15080,"tokens_out":3379,"duration_ms":34134,"significance":"If the claims hold, the RMB approach is a valuable contribution to sparse STEM-EDX analysis: it provides denoised quantitative maps with per-pixel uncertainty, is applicable at extremely low counts where matrix-decomposition methods struggle, and is backed by publicly available code on GitHub. The use of physics-based ESPM endmembers and the multiscale spatial prior is a sensible combination, and the independent HAADF-based column-location comparison provides a strong spatial-resolution benchmark for the experimental data. However, the absolute accuracy claim rests on simulations generated with the same ESPM forward model used for quantification, which tests inversion consistency rather than physical fidelity, and the surrogate likelihood fixes initial abundances from the same data, so the reported uncertainties do not account for that estimation step. These limitations need to be addressed or clearly bounded before the accuracy claims can be fully accepted.","major_comments":[{"comment":"Equation (2) replaces the Poisson likelihood with a gamma surrogate whose shape parameter is 1 + r̅_{n,k}, where r̅ is the SUNSAL estimate obtained from the same spectrum image. The data therefore enter twice: once to set the effective photon count in the likelihood and once in the observed counts themselves. This is load-bearing because the reported posterior variances ψ_{n,k} are computed conditional on this fixed r̅ and do not propagate the uncertainty in the SUNSAL initialization or the approximation error of Jensen's inequality. I recommend calibrating the uncertainty output on the simulated data: check whether the 68% or 95% credible intervals derived from ψ contain the known ground-truth abundances at the expected rates, and test sensitivity of the maps and uncertainties to perturbed or alternative initial abundance estimates. Without such a check, the uncertainty maps should be described as conditional rather than fully calibrated posterior uncertainties.","section":"Methods, Algorithm description; SI, Derivation of eq. (2)"},{"comment":"The central accuracy claim—that the chemical maps are 'more accurate' than standard methods—is supported on simulated data by inverting the same ESPM forward model used to generate the spectra. This is an internal-consistency test: it demonstrates that the algorithm can recover known abundances from its own forward model, but it does not validate the physical fidelity of the ESPM cross-sections, Bremsstrahlung background, or detector response for the experimental geometry. The experimental section itself states that atomic-resolution EDX concentrations are 'not always directly meaningful, owing to electron channelling', and the simulated sample is spatially uniform rather than atomic-resolution. I recommend either providing an external accuracy anchor (for example, comparison with EELS quantification, a stoichiometric standard, or a thickness series where absorption can be estimated) or explicitly limiting the accuracy claim to 'precision and spatial resolution' while stating that absolute concentrations inherit the validity of the ESPM model.","section":"Data simulation; Results, Simulated Data; Results, Experimental Data"},{"comment":"The Methods state that the ESPM-based endmember approach is valid 'for relatively thin samples, where absorption does not represent a relevant contribution', but the experimental PbTiO3 film is 28 nm thick and is imaged at atomic resolution, where electron channelling is acknowledged to complicate quantification. The magnitude of X-ray absorption and channelling effects at this thickness and for the Pb L and Ti K lines is not estimated or discussed. This is load-bearing for the quantitative elemental maps, although the unit-cell c/a demonstration is less affected because it relies on relative column positions. Please quantify or justify the thin-sample approximation for the specific experimental conditions, or soften the quantitative claims accordingly.","section":"Methods, Algorithm description; Experimental data acquisition"}],"minor_comments":[{"comment":"The SUNSAL algorithm (Ref. 24) is typeset inconsistently as 'Sunsal' in the main text and supplementary information; use 'SUNSAL' consistently.","section":"Methods, Algorithm description"},{"comment":"The inverse-gamma hyperparameters α_l and β_l are said to be 'set to zero', which would make the prior improper; please clarify whether this is intended as a limiting non-informative prior and how the implementation handles it.","section":"Methods, Algorithm description, Eq. (5)"},{"comment":"The notation for multiscale kernel sizes is ambiguous: the text refers to '2l-1 by 2l-1' windows and later to downsampling sets {1, 3, 5, …, 2l+1}; please define the relationship between the level index l and the kernel size unambiguously.","section":"Methods, Algorithm description, Eq. (3) and Results, Simulated Data"},{"comment":"The quantity labeled 'SNR ≡ σ/I' is actually a noise-to-signal ratio (lower is better); the label and the discussion should be inverted or renamed to avoid confusion.","section":"Results, Simulated Data, Fig. 4"},{"comment":"In the paragraph describing Figure 6, the sentence 'all the atomic columns are correctly identified in the RMB map correctly' contains a duplicated adverb; please revise.","section":"Results, Experimental Data"}],"recommendation":"major_revision","confidential_remarks":"The paper's own stated limitations—the thin-sample assumption for ESPM and the channelling caveat for atomic-resolution EDX—are in tension with the unqualified accuracy claim in the abstract. The internal-consistency nature of the simulation benchmark and the conditional nature of the uncertainties should be prominently acknowledged. The spatial-resolution improvements on experimental data are convincing and valuable, and the code availability is a plus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Pau, this is a solid methods paper and a fair one. The core idea—marrying the multiscale Bayesian Poisson reconstruction that Halimi and others developed for single-photon LIDAR with ESPM-simulated EDX endmembers—is genuinely new and well implemented. The paper does what a good methods paper should: it gives a clear statistical model, a coordinate-descent algorithm with analytical updates, code on GitHub, and benchmarks against synthetic data with known ground truth at 5, 50, and 500 counts per pixel. The comparison to standard Cliff-Lorimer quantification is fair, since both methods get the same spatial smoothing. The experimental PbTiO3 demo uses HAADF images as an independent reference for atomic column positions and for the c/a ratio maps, which is the right way to ground the spatial-resolution claim.\n\nThe stress-test worry about the ESPM self-consistency loop is real but not as damaging as it sounds. The simulated data are generated with ESPM and then inverted with ESPM endmembers, so the quantitative accuracy test is internal-consistency. That is standard practice in this community, but it does mean the claim of 'more accurate chemical maps' shouldn't be read as externally validated absolute concentrations—the paper's own experimental section wisely says atomic-resolution EDX concentrations are not always meaningful due to channelling. The abstract and conclusions lean a bit harder on 'more accurate' than the evidence supports; the experimental evidence is about spatial resolution and structural features, not unbiased abundances. That's a mild overclaim, not a fatal one.\n\nThe more specific technical soft spot is the surrogate likelihood in Eq. (2), which fixes r_bar using Sunsal on the same dataset. The data therefore enter twice, and the reported per-pixel uncertainties don't include the error in those initial estimates. I'd like to see a sensitivity analysis or a simulation-based calibration check for the uncertainty maps. The paper also lacks repeated-simulation error bars on the SNR/broadening curves in Fig. 4; each point appears to come from a single synthetic dataset. That's easy to fix and would make the comparison much more convincing.\n\nOverall, this deserves a serious referee. The method is reproducible, the evaluation is thoughtful, and the experimental demonstration is honest about its limitations. I'd recommend conditional acceptance: ask for the Sunsal sensitivity check, error bars on the benchmark plots, and a slightly softer phrasing of the accuracy claim in the abstract.","headline":"A well-executed and reproducible method paper for sparse STEM-EDX quantification; the main caveat is that quantitative accuracy is only demonstrated on simulations that use the same physics library the method is built on, though the experimental section stays on the spatial-resolution side.","tokens_in":15617,"tokens_out":2747,"would_cite":true,"duration_ms":26276,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Poisson-aware multiscale Bayesian estimator recovers accurate, denoised elemental maps from sparse STEM-EDX spectra, resolving ferroelectric domains at unit-cell scale in PbTiO3.","keywords":["Bayesian estimation","EDX spectrum imaging","STEM","chemical quantification","denoising","Poisson statistics","multiscale","ferroelectric domains"],"falsifier":"Acquire an EDX spectrum image from a wedge-shaped standard specimen of known composition whose thickness varies from about 10 nm to 100 nm, run the RMB algorithm, and check whether the recovered atomic percentages drift systematically with thickness or deviate from known values in thicker regions; any such drift would show that the simulated endmembers do not capture absorption effects.","tokens_in":14666,"feed_emoji":"🔬","tokens_out":5554,"duration_ms":52325,"temperature":0.7,"pith_summary":"Energy-dispersive X-ray (EDX) spectrum imaging in a scanning transmission electron microscope is a powerful but photon-starved technique: each pixel often records only a few counts, and standard quantification then produces noisy, unreliable maps. The authors introduce a robust multiscale Bayesian (RMB) estimator that models the detected counts as Poisson, uses simulated elemental and Bremsstrahlung spectra as endmembers, and borrows statistical strength from neighbouring pixels at several spatial scales. They claim that on simulated data this yields more accurate chemical maps at higher spatial resolution than standard line-integration and Cliff-Lorimer quantification, while also returning per-pixel uncertainty. On an experimental atomic-resolution PbTiO3 dataset, the RMB maps resolve ferroelectric domains with unit-cell resolution, something the standard analysis cannot do at the same acquisition time. The broader payoff is a general, training-free pipeline that extracts near-limit chemical information from beam-sensitive or short-acquisition EDX measurements.","feed_headline":"Bayesian method sharpens EDX chemical maps to unit-cell resolution","feed_subtitle":"Poisson-aware multiscale estimation extracts accurate compositions and uncertainty from sparse, noisy X-ray spectra.","key_machinery":"The central object is a hierarchical Bayesian model with a Poisson-Gamma observation model: after a Jensen-inequality approximation, the per-pixel likelihood becomes a product of Gamma distributions over abundances, with rate terms $S_k$ from simulated endmembers. Spatial correlation enters through a latent variable $m_{n,k}$ with a Gaussian Markov-random-field prior whose weights are exponential functions of abundance differences at multiple scales; inverse-Gamma priors on the variances $\\psi_{n,k}$ yield uncertainty estimates. Point estimates come from maximum a posteriori optimisation by coordinate descent, with analytical updates for the abundance maps $r_{n,k}^{(\\ell)}$, the latent map $m_{n,k}$, and the variance $\\psi_{n,k}$. Initial abundances are obtained from Sunsal unmixing, and the endmembers themselves come from the ESPM library, which accounts for acceleration voltage, take-off angle, and detector efficiency.","core_discovery":"The central claim is that a fully Bayesian treatment of EDX spectrum images, in which each pixel's spectrum is generated by a Poisson mixture of simulated atomic-emission and Bremsstrahlung endmembers and abundances are spatially regularised through a latent Gaussian field evaluated at multiple resolutions, outperforms standard quantification. Specifically, the paper argues that the recovered elemental maps are more accurate and preserve higher spatial resolution than conventional kernel-smoothed line-integration/Cliff-Lorimer maps at equal signal levels. The authors demonstrate this on simulated PbTiO3/SrRuO3 spectrum images with 5, 50, and 500 counts per pixel, and then on a 28 nm PbTiO3 film where the RMB maps of Pb and Ti show sharper atomic columns and additional Fourier spots, allowing atomic columns to be located like HAADF and unit-cell c/a ratios to be measured across a ferroelectric domain boundary.","pith_inferences":["Extension: the same Poisson-multiscale-Bayesian machinery could be reused for other counting detectors, such as EELS, time-of-flight, or X-ray ptychography, wherever known spectral responses and spatial smoothness hold.","Extension: the uncertainty maps suggest an adaptive version of the algorithm that chooses per-pixel kernel sizes by minimising predicted variance, since large kernels degrade interface localisation.","Extension: if electron-channelling and absorption were incorporated into the endmembers, the method could go beyond thin specimens and make atomic-resolution concentrations physically meaningful, a step the paper does not take.","Extension: a thickness series on a known specimen would directly test the method's boundary, because simulated endmembers ignore absorption."],"forward_implications":["RMB quantification stays accurate down to roughly 5 X-ray counts per pixel spectrum, a regime where matrix-decomposition methods cannot operate.","For a fixed kernel size, RMB maps show higher signal-to-noise ratio and less feature broadening than standard kernel-smoothed Cliff-Lorimer maps on simulated data.","Per-pixel uncertainty estimates come directly from the Bayesian posterior, without propagating k-factor, background, and counting errors separately.","Experimental PbTiO3 maps from only 350 s of acquisition locate all atomic columns correctly relative to HAADF, enabling unit-cell c/a ratio maps that match HAADF-derived values.","Because the posterior can be updated with new data, the algorithm is compatible with on-the-fly abundance estimation during live acquisition."],"supporting_citations":[{"why":"Supplies the simulated elemental X-ray emission spectra and Bremsstrahlung background endmembers used in the observation model.","marker":"[19]"},{"why":"Provides the robust Bayesian reconstruction framework originally developed for single-photon LIDAR data that the algorithm builds on.","marker":"[23]"},{"why":"The Sunsal algorithm generates the initial abundance estimates used to start the coordinate descent optimisation.","marker":"[24]"},{"why":"Defines the standard Cliff-Lorimer quantification approach against which RMB results are compared.","marker":"[2]"},{"why":"HyperSpy routines implement the standard line-integration and background-fitting workflow used as the comparison baseline.","marker":"[35]"},{"why":"Atomap locates atomic columns in the EDX and HAADF images, enabling quantitative comparison of spatial resolution.","marker":"[43]"},{"why":"Connects the measured unit-cell distortion (c/a ratio) to ferroelectric domain structure in c-phase lead titanate thin films.","marker":"[42]"}],"fun_headline_variants":["Bayesian EDX analysis maps atoms with unit-cell precision","Poisson-aware Bayesian maps sharpen X-ray chemical imaging","Multiscale Bayesian denoising boosts EDX atomic maps","New Bayesian tool extracts sharper chemical maps from sparse X-rays","Unit-cell resolution EDX maps via Bayesian multiscale model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline depends on the simulated endmember spectra correctly matching the real X-ray emission and background of the specimen, and the paper itself limits this to thin samples where absorption is negligible while noting that electron channelling makes atomic-level concentrations not directly meaningful.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian EDX analysis maps atoms with unit-cell precision","Poisson-aware Bayesian maps sharpen X-ray chemical imaging","Multiscale Bayesian denoising boosts EDX atomic maps","New Bayesian tool extracts sharper chemical maps from sparse X-rays","Unit-cell resolution EDX maps via Bayesian multiscale model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1182,"prompt_tokens":892,"completion_tokens":290,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":208}},"tokens_in":508,"tokens_out":290,"duration_ms":3462,"temperature":1.0,"reasoning_tokens":208,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T12:36:43.946019+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Acquire an EDX spectrum image from a wedge-shaped standard specimen of known composition whose thickness varies from about 10 nm to 100 nm, run the RMB algorithm, and check whether the recovered atomic percentages drift systematically with thickness or deviate from known values in thicker regions; any such drift would show that the simulated endmembers do not capture absorption effects.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the simulated elemental X-ray emission spectra and Bremsstrahlung background endmembers used in the observation model."},{"cited_title":"A., Buller, G","cited_arxiv_id":null,"evidence_quote":"Provides the robust Bayesian reconstruction framework originally developed for single-photon LIDAR data that the algorithm builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Sunsal algorithm generates the initial abundance estimates used to start the coordinate descent optimisation."},{"cited_title":"& Lorimer, G","cited_arxiv_id":null,"evidence_quote":"Defines the standard Cliff-Lorimer quantification approach against which RMB results are compared."},{"cited_title":"de la et al","cited_arxiv_id":null,"evidence_quote":"HyperSpy routines implement the standard line-integration and background-fitting workflow used as the comparison baseline."},{"cited_title":"E., MacLaren, I., Tybell, T","cited_arxiv_id":null,"evidence_quote":"Atomap locates atomic columns in the EDX and HAADF images, enabling quantitative comparison of spatial resolution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Connects the measured unit-cell distortion (c/a ratio) to ferroelectric domain structure in c-phase lead titanate thin films."}],"review_version":1}