{"id":"73c7e496-6185-4208-afd6-2d28552f1b9c","arxiv_id":"1908.07773","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"SITES is a fast, non-subjective iterative solver that recovers coronal differential emission measures from AIA EUV images and performs well on synthetic tests.","lead":"This paper introduces SITES, a new algorithm that estimates the temperature distribution of the Sun's corona from EUV images. It is simple, fast, and performs comparably to existing methods on synthetic tests, while honestly documenting its limits for narrow and cool plasma.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation 4 as printed does not preserve channel intensities: the per-channel term is I_i/(R_ij ΔT_j) times a sum of response squares, so summing D_j R_ij ΔT_j gives n_t I_i A_i, not I_i; the central algorithm is therefore not the one validated in §3.","rationale":"Read in good faith: the paper intends to present a simple iterative DEM inversion, and the central claim is that SITES accurately recovers model DEMs from AIA synthetic data, compares well with SMI/TR, and runs at about 1000 DEMs per second. For that claim to hold, Eq. 4 must define an intensity-preserving initial and residual redistribution that a reader can implement. The provided text fails this condition: the printed per-channel term is the inverse of the usual least-norm backprojection, as shown by the integration and dimensional checks. This is not a tuning detail or a limitation statement; it is the exact operation feeding the iteration. Consequently the numerical sections, however plausible, do not verify the algorithm as written. I also note secondary issues the reader already identified: the smoothing width is chosen by trial and error, and real-data accuracy inherits AIA response-function errors; those would only warrant conditional acceptance. The Eq. 4 issue is more serious because it makes the manuscript's central algorithm unreproducible as written, so I recommend UNVERDICTED until the equation is corrected and, ideally, code or a worked numerical example is provided. I disagree with the reader's choice of weakest assumption: response-function bias is a standard limitation of all DEM methods using AIA responses, while the Eq. 4 inconsistency is internal to the method's definition and would invalidate the reported validation if it is not a transcription error.","tokens_in":13967,"tokens_out":10206,"duration_ms":104187,"concrete_test":"Implement Eq. 4 exactly as printed, using the stated 43 log-temperature bins spanning 0.07-20 MK, Gaussian smoothing with σ=3.2 bins, positivity truncation each iteration, and the 1% convergence criterion. Run the §3.1 single-Gaussian target DEM and record the per-iteration TI residual and the output DEM. Also compute the per-channel integrals Σ_j D_j R_ij ΔT_j for the initial DEM. If the integrals differ from I_i by a factor of n_t Σ_j (R_ij ΔT_j)^2, or the iteration diverges or never reaches the 1% threshold, then Eqs. 4-5 as printed do not reproduce Figs. 2-13 and the algorithm description must be corrected before the central claim can be evaluated.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"Section 2, Eq. 4 is the core of the method, but as printed it cannot be the expression used to produce the paper's results. The bracketed per-channel term is I_i/(R_ij ΔT_j) * Σ_j (R_ij ΔT_j)^2. For a single channel, summing D_j R_ij ΔT_j over the n_t temperature bins gives n_t I_i Σ_j (R_ij ΔT_j)^2, not I_i. The expression is also dimensionally wrong: it carries units of intensity times response, not DEM. The standard intensity-preserving backprojection would be I_i R_ij ΔT_j / Σ_j (R_ij ΔT_j)^2. If the printed form is taken literally, the update diverges wherever R_ij approaches zero (e.g., 94/304 channels at low temperatures), and the claimed convergence, the §3 synthetic recoveries, the §3.5 method comparison, and the ~1000 DEM/s timing cannot follow from the stated algorithm. No derivation or public code is provided, so the text alone does not specify a valid, reproducible method.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents SITES, an iterative differential emission measure (DEM) inversion method for AIA EUV observations. The method builds an initial DEM by backprojecting each channel's intensity across temperature according to the channel response, then iteratively updates the DEM using intensity residuals with positivity and smoothing constraints. The authors test SITES on synthetic DEMs (single and multi-Gaussian, with and without noise), measure its speed, compare it with the methods of Cheung et al. (2015) and Hannah & Kontar (2012), and apply it to full-disk AIA data. They also introduce a fractional emission measure (FEM) visualization. The central claims are that SITES is simple, fast (about 1000 DEMs per second), non-subjective apart from one smoothing parameter, and accurate for broad DEMs above ~0.5 MK, while performing poorly on narrow peaks and low-temperature DEMs.","tokens_in":14183,"tokens_out":5574,"duration_ms":52301,"significance":"If the method works as described, it is a useful addition to the DEM inversion toolbox: the synthetic validation is extensive, the authors explicitly characterize the failure modes (narrow peaks, temperatures below ~0.5 MK), and the uncertainty estimate is checked against Monte Carlo repetitions. The FEM visualization is a simple and effective idea for comparing temperature regimes across coronal structures. The speed and simplicity could make it practical for large AIA datasets. However, the central equation of the method as printed is not intensity-preserving, so the validation presented in Section 3 cannot be reproduced from the manuscript as written; this is a load-bearing issue that must be corrected before the results can be credited.","major_comments":[{"comment":"As printed, Eq. (4) cannot be the backprojection used in the paper. The per-channel term is I_i/(R_ij ΔT_j) times Σ_j (R_ij ΔT_j)^2, so for a single channel the forward-modeled intensity obtained by summing D_j R_ij ΔT_j over j is n_t I_i Σ_j (R_ij ΔT_j)^2, not I_i. The expression is also dimensionally inconsistent: it carries units of intensity times response, not DEM. The standard intensity-preserving backprojection would be I_i R_ij ΔT_j / Σ_j (R_ij ΔT_j)^2, with the sum in the denominator. Because no derivation or public code is provided, the text alone does not specify a valid, reproducible algorithm, and the convergence, timing, and comparison results in Section 3 cannot be traced to the stated equation. Please correct Eq. (4), provide a derivation, and state explicitly which expression was used in the synthetic tests.","section":"Section 2, Eq. (4)"},{"comment":"The quantity d_j defined in Eq. (6) is dimensionless: it is the square root of a weighted sum of squared relative errors (σ_i/I_i)^2 and ε_i^2. Yet it is presented as 'the final DEM uncertainty' and is compared in Figures 2, 9, and 10 to the absolute spread of DEMs, which has units cm^-5 K^-1. If d_j is intended as a relative uncertainty, this must be stated and the plotted error bars must be multiplied by the DEM value; if it is intended as an absolute error, the equation is incomplete. This issue directly affects the validation claim in Section 3.4 that the uncertainty estimate 'reflects well the true variation of the output DEMs.'","section":"Section 2, Eq. (6)"},{"comment":"The claim that SITES is 'non-subjective' and that the smoothing kernel width is the only parameter affecting the result is overstated. The kernel width is chosen by trial and error with the criterion that it be minimal 'whilst still resulting in smooth DEMs', which is a subjective regularization choice. Moreover, Section 3.3 shows that the convergence threshold also affects both the fitted intensities and the output DEM (Figure 8a), so the convergence threshold is another parameter that influences the result. The non-subjectivity claim should be qualified to 'a single tunable regularization parameter plus a stopping criterion', and the dependence on the convergence threshold should be acknowledged in the summary.","section":"Sections 1 and 5"},{"comment":"The text in Section 3.3 refers to a 'complex 3-Gaussian plus background' DEM, while the complex test in Section 3.2 is explicitly defined as two Gaussian peaks plus a constant background. This inconsistency makes it unclear which model was used for the speed and convergence-threshold tests. Please correct the description so that the number of Gaussians matches the actual test DEM.","section":"Section 3.3"}],"minor_comments":[{"comment":"There is a typo: 'perofrmance' should be 'performance'.","section":"Section 3.5"},{"comment":"There is a typo: 'measurmeent' should be 'measurement'.","section":"Section 4.1"},{"comment":"The sentence following Eq. (4) says that integrating the individual DEMs over temperature would result in exactly the observed intensities; this statement is only true for the corrected form of the equation, not for the printed form. Please move this explanatory sentence next to the corrected equation.","section":"Section 2"},{"comment":"The software is only available 'by email request to the authors'. For reproducibility, especially given the error in Eq. (4), please consider providing a version-controlled repository or a complete pseudocode listing that includes the exact update rule and convergence criterion.","section":"General"},{"comment":"In the comparison with SMI, the plots show EM rather than DEM for the SMI output, and the text notes this in the caption; however, the y-axis labels are not visible in Figures 11a and 11b. Please ensure the figure axes make the EM/DEM distinction explicit.","section":"Section 3.5"},{"comment":"The abstract states 'about 1000 DEMs per second', but Section 3.3 reports this speed at a convergence threshold of 4%. The threshold dependence is shown in Figure 8b, so the abstract should either quote the speed with the associated threshold or mention that the speed is threshold-dependent.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The discrepancy in Eq. (4) is likely a typesetting error, but because no code is provided, I cannot determine the actual algorithm from the manuscript. The authors should be required to supply the corrected equation, a derivation, and ideally the code or a precise pseudocode version, so that the synthetic validation in Section 3 can be checked against the stated method. The paper otherwise fits the scope of Solar Physics and the synthetic test design is a strength, but the central algorithm description must be made correct and reproducible before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful method paper, but the central equation is garbled. I checked the stress-test concern and it lands. In Eq. 4, the per-channel term is I_i/(R_ij ΔT_j) times Σ_j (R_ij ΔT_j)^2. Integrate that DEM against the response and you get n_t I_i times that sum, not I_i. The text says each channel's DEM reproduces the observed intensity by construction, so either the equation or the implementation is wrong. Since the §3 synthetic recoveries and the comparisons in §3.5 are clean and plausible, my guess is a typesetting error—probably the intended expression is I_i R_ij ΔT_j / Σ_j (R_ij ΔT_j)^2. But the manuscript as submitted cannot be used to reproduce the method.\n\nWhat is genuinely good: the synthetic testing is careful. They use known DEMs, realistic AIA responses, and report multiple metrics (correlation, intensity residual, DEM deviation). They explicitly map where the method fails—narrow peaks and cool plasma—and that matches the known ill-posedness of the problem. The comparison with Cheung et al. and Hannah & Kontar is honest, including the note that SMI is much faster. The FEM visualization in §4 is a simple and effective idea, and the application to full-disk AIA data looks sensible.\n\nThe soft spots, in proportion: the Eq. 4 problem is the big one. It is load-bearing, not a minor typo; without the correct equation, the reader cannot know what was actually run. The lack of public code—only \"available by email request\"—makes this worse. I'd want the code posted or a supplement with the corrected equation. The smoothing kernel width is found by trial and error; the authors say this is the only subjective choice, which is fair, but it means the method has one tunable knob plus the convergence threshold. Not a flaw, just worth knowing.\n\nBottom line: the paper is for solar physicists who process AIA DEMs and want something simple, fast, and less subjective than regularized inversion. The method is likely sound, but the text is not. This deserves a serious referee, but the authors need to correct Eq. 4 and make the code available before it is publishable.","headline":"A well-designed DEM solver with honest tests, but the printed Eq. 4 is wrong—the paper's own results cannot follow from the stated algorithm without a fix and released code.","tokens_in":14701,"tokens_out":4635,"would_cite":false,"duration_ms":43203,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SITES recovers coronal temperature structure from EUV images using a simple iterative solver.","keywords":["differential emission measure","solar corona","EUV imaging","AIA/SDO","temperature response functions","iterative inversion","fractional emission measure","optically thin plasma"],"falsifier":"A blind benchmark on synthetic AIA data spanning central temperatures from log T = 5.7 to 7.0 and widths from log T = 0.1 to 0.35, with Poisson noise added, would settle the accuracy claim: the central claim is wrong if SITES systematically shows input–output DEM correlations below 95% inside the region the paper marks as reliable, or if its measured residuals do not fall below the convergence threshold it reports.","tokens_in":13736,"feed_emoji":"☀️","tokens_out":6554,"duration_ms":91963,"temperature":0.7,"pith_summary":"Solar EUV images of the optically thin corona carry temperature information only indirectly: each channel mixes emission from many temperatures. This paper presents SITES, a differential emission measure (DEM) inversion that turns several EUV channel intensities into a temperature profile of the emitting plasma. The paper aims to show that a direct, weighted redistribution of observed intensities according to each channel's temperature response, refined by iterating on residuals, can recover model DEMs from synthetic AIA observations with accuracy comparable to established methods. It also shows the solver is fast enough for full-disk, long-duration datasets and introduces fractional emission measure maps that make temperature dominance in coronal structures visible at a glance.","feed_headline":"SITES: 1,000 coronal temperature profiles per second from EUV images","feed_subtitle":"A one-parameter iterative solver matches established DEM inversions on synthetic AIA data and maps where each temperature dominates.","key_machinery":"The core object is the relative temperature response $S_{ij} = w_i R_{ij} / (\\sum_i w_i R_{ij})$, built from each channel's response function $R_{ij}$ weighted by measurement and calibration errors. Equation (4) converts observed intensities $I_i$ into an initial DEM by distributing each channel's intensity across temperature according to $R_{ij}$, combining channels through $S_{ij}$, and convolving with a narrow Gaussian in logarithmic temperature. The iteration then computes modelled intensities $M_i = \\sum_j D_j R_{ij} \\Delta T_j$, feeds the residual $I_i - M_i$ back through the same redistribution, and clips negative DEM values to zero, stopping when the weighted mean absolute residual falls below a threshold such as 4%. This iterative residual-correction loop, together with the single smoothing width, is what carries the whole argument.","core_discovery":"The central claim is that SITES reconstructs a DEM by first spreading each measured intensity across temperature in proportion to that channel's response function, combining the channels with inverse-noise weights, then repeatedly adding DEM corrections built from the residual intensities until the weighted residuals drop below a threshold. In tests on synthetic AIA observations, the recovered DEMs correlate at 95–98% with the input model across broad ranges of peak temperature and width, with median deviations around 12–26% for representative single- and double-Gaussian targets. The method is deliberately minimal: the only user-selected parameter is the width of a Gaussian smoothing kernel in logarithmic temperature, with positivity enforced by thresholding. The paper concedes two limits—very narrow DEM peaks are not well recovered at any temperature, and temperatures below about 0.5 MK are not reliable with AIA data—and argues these limits follow from the instrument's broad response functions rather than from the solver. On identical synthetic inputs, SITES matches or outperforms both a sparse-basis method and a Tikhonov-regularized method used as benchmarks, while running at roughly a thousand DEMs per second at the routinely used convergence threshold.","pith_inferences":["A direct extension is to feed SITES DEMs into rotational tomography: using multiple viewpoints would address the line-of-sight bias the paper notes for off-limb regions, and could also regularize narrow peaks.","The error bars from Equation (6) omit smoothing and iteration effects; the paper's own low-signal test shows underestimated uncertainty above ~2 MK, so users should treat those error bars as optimistic in faint regions.","The real 0.5 MK floor is a property of AIA's response functions, not of the algorithm; applying SITES to observations with stronger low-temperature sensitivity would likely push the useful range lower, a testable prediction.","Because the iteration is a residual-correction map with positivity clipping, it should be possible to prove convergence conditions, which might let future users choose thresholds automatically rather than by trial."],"forward_implications":["Long time-series of AIA full-disk DEM maps become practical: at about 1,000 profiles per second, a full 4k disk can be processed in minutes to hours rather than days.","Multi-channel EUV instruments beyond AIA can use the same solver directly, provided their temperature response functions and error estimates are known.","The fractional emission measure visualization gives a standard, normalized way to compare which temperatures dominate in coronal holes, quiet Sun, and active region cores, independent of total emission.","The method's reliance on a single smoothing parameter makes it easier to reproduce and less dependent on analyst choices than methods with several tuning parameters.","Broad multithermal coronal structures are the natural target; narrow isothermal-like plasmas and sub-0.5 MK material should be interpreted with caution."],"supporting_citations":[{"why":"Supplies the CHIANTI atomic data used to compute the AIA channel temperature response functions on which the inversion rests.","marker":"Dere et al., 1997"},{"why":"Provides updated CHIANTI version 7 atomic data used in the same response functions.","marker":"Landi et al., 2012"},{"why":"Provides the EVE cross-calibration and the 94 Å correction built into the response functions.","marker":"Boerner et al., 2014"},{"why":"Serves as one of the established benchmark methods in the synthetic-data comparisons.","marker":"Cheung et al., 2015"},{"why":"Serves as the Tikhonov-regularization benchmark and the source of the overview of DEM inversion methods.","marker":"Hannah and Kontar, 2012"},{"why":"Provides the earlier residual-iteration DEM approach that SITES resembles in its iterative correction of negative intensities.","marker":"Plowman, Kankelborg, and Martens, 2013"}],"fun_headline_variants":["SITES: Solar DEM inversion at 1,000 profiles per second","SITES: Minimal assumptions, maximal speed for DEM maps","SITES: Fast and simple DEM reconstructions from EUV data","SITES: 1000 DEMs/sec, matches benchmark methods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reconstruction inherits whatever bias lives in the AIA temperature response functions; if those functions are wrong, the iterative fitting cannot correct it, because the method only tries to match intensities computed with the same functions.","fun_headline_variants_meta":{"raw":{"variants":["SITES: Solar DEM inversion at 1,000 profiles per second","SITES: Minimal assumptions, maximal speed for DEM maps","SITES: Fast and simple DEM reconstructions from EUV data","SITES: 1000 DEMs/sec, matches benchmark methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000393,"raw_usage":{"total_tokens":2134,"prompt_tokens":1087,"completion_tokens":1047,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":703,"completion_tokens_details":{"reasoning_tokens":973}},"tokens_in":703,"tokens_out":1047,"duration_ms":33995,"temperature":1.0,"reasoning_tokens":973,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:55:55.488359+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A blind benchmark on synthetic AIA data spanning central temperatures from log T = 5.7 to 7.0 and widths from log T = 0.1 to 0.35, with Poisson noise added, would settle the accuracy claim: the central claim is wrong if SITES systematically shows input–output DEM correlations below 95% inside the region the paper marks as reliable, or if its measured residuals do not fall below the convergence threshold it reports.","supporting_citations":[],"review_version":1}