{"id":"7081eefb-050e-4314-a772-616dd4ed7f63","arxiv_id":"2507.03640","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A ptychography reconstruction framework where the cropping shift of each diffraction pattern is optimized with automatic differentiation, recovering subpixel misalignments without manual tuning.","lead":"This paper adds a trainable diffraction-pattern shift parameter to automatic-differentiation ptychographic reconstruction, aiming to correct crop misalignment automatically. Simulations show subpixel recovery of offsets up to 5 pixels, and EUV experiments yield sharper reconstructions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The <0.5 px subpixel-accuracy claim is validated only on data generated with the same shift operator used in reconstruction, so real-world shift model mismatch is untested.","rationale":"The reader's weakest_assumption and my concern coincide: the simulation's matched-model design is the most load-bearing weakness. I reviewed Algorithm 1 and Eq. 4: no mathematical inconsistency is apparent, and the idea of optimizing ξ with AD is sound and straightforward. But the abstract's quantitative claim ('average deviation below 0.5 pixels') is made on the basis of §3.1, and §3.2 has no quantitative shift ground truth. If the concern lands, the paper's central claim is overgeneralized from an inverse-crime simulation; if it does not land, e.g., an out-of-model test still yields <0.5 px, the claim is solid. The appropriate verdict therefore remains CONDITIONAL, matching the reader. I do not escalate to REJECT because the method is not circular and the experimental improvement is at least suggestive. The main missing piece is a test where the ground-truth shifts are produced outside the reconstructor's shift manifold.","tokens_in":11518,"tokens_out":7623,"duration_ms":99813,"concrete_test":"Modify the simulation so the ground-truth measured patterns are generated with a shift operation different from Algorithm 1's f: compute each diffraction pattern on a fine grid, apply a continuous subpixel shift by Fourier interpolation or by high-accuracy resampling, then integrate over 2×2 or 3×3 detector pixel bins to emulate pixel response; reconstruct with the original f and report the mean |ξ_recovered − ξ_true| for a=5. Repeat with integer-pixel shifts followed by downsampling and with a different interpolation kernel (e.g., bicubic vs bilinear). If mean error stays below 0.5 px in these out-of-model tests, the inverse-crime concern is resolved; if not, the subpixel claim is conditional on the exact shift model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 reports mean shift errors below 0.5 px for offsets up to 5 px, but the simulation appears to generate the ground-truth shifts with the same differentiable operator f(·,ξ) that Algorithm 1 line 5 optimizes. The text never states the interpolation/resampling used for the ground-truth offsets; if it is the same as f, the test is an inverse-crime validation: the optimizer only needs to invert the exact model that produced the data. Real EUV detector misalignment is not guaranteed to lie in that model family: pixel-area integration, the detector point-spread function, the particular interpolation kernel used in TPC, and residual geometric distortions all alter how a true physical offset appears in the measured pixels. Under such mismatch, optimizing ξ will reduce the loss but the recovered shift need not equal the true crop offset, and the 0.5 px bound has no demonstrated transfer. The experimental section (§3.2) does not provide ground-truth shifts; it only shows improved fringe contrast when ξ is optimized, which could also result from the additional degrees of freedom or from joint optimization of δ, φ, and I_b. Therefore the paper's central quantitative claim is currently supported only by a matched-model simulation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an automatic-differentiation-based ptychographic reconstruction framework in which the cropping shift of each recorded diffraction pattern is treated as a learnable parameter. Algorithm 1 embeds the shift variable xi in the forward model through an operator f, alongside the object, probe, scan-position deviations, tilt angles, and background term. The authors validate the approach with simulations of reflection, broadband, and reflection-broadband models with preset offsets up to 5 pixels, reporting mean shift deviations below 0.5 pixels, and with an EUV reflection experiment in which optimizing the shift variables visibly sharpens reconstructed fringes. The paper claims that the method achieves subpixel correction without manual tuning and is modular across ptychographic modalities.","tokens_in":1360,"tokens_out":1916,"duration_ms":110024,"significance":"If the quantitative claim is established, the paper is a useful incremental contribution to AD-based ptychography: it removes a manual preprocessing step, is formulated in a modular way that can be combined with tilted-plane correction and broadband models, and the experimental section shows a clear visual improvement in an EUV reflection setup. The authors are also transparent about known sources of residual error, such as anisotropic resolution in the reflection model and information loss in numerical monochromatization. The main weakness is that the central subpixel-accuracy claim rests on a matched-model simulation in which the ground-truth shifts may be generated by the same differentiable operator that is optimized, and the experimental section provides no ground-truth shift values. The paper therefore does not yet establish that the method is accurate under real detector misalignment, which is the load-bearing claim of the manuscript.","major_comments":[{"comment":"The central quantitative claim of an average shift error below 0.5 pixels is supported only by a matched-model simulation. The text does not state how the ground-truth subpixel offsets are generated in Section 3.1; if they are applied with the same differentiable operator f(., xi) that Algorithm 1 line 5 optimizes, then the simulation is an inverse-crime test that verifies the optimizer can invert its own forward model but not that real detector misalignments obey that translation model. Real misalignments involve pixel-area integration, detector point-spread function, the interpolation kernel used in TPC, and residual geometric distortion. Please specify the interpolation or resampling used for the ground-truth offsets and add mismatch experiments, for example generating shifts with a different kernel, including pixelation and Poisson noise, or applying a small affine distortion, to show that the claimed subpixel accuracy transfers to model mismatch.","section":"Section 3.1 and Algorithm 1, line 5"},{"comment":"The experimental section does not provide ground-truth shift values, so it cannot independently validate the subpixel accuracy claim. The comparison is between reconstruction with and without AD shift correction, not against a conventional centering method, and the observed improvement in fringe contrast could partly result from the extra optimizable degrees of freedom, especially in the joint-optimization case in Fig. 6(c). Please include a comparison with a standard centering baseline, such as zero-order centroid or cross-correlation alignment, and, if possible, report the learned shift parameters and their consistency across independent reconstructions, or inject a known subpixel shift into a measured pattern to obtain a quantitative experimental error.","section":"Section 3.2, Fig. 6"},{"comment":"The simulation protocol does not include noise or model uncertainty, and the method depends on several hyperparameters that are not tested for sensitivity, including the object learning rate of 0.1, the offset learning rate of 0.95, the 5 warmup epochs without offset correction, and the minibatch size of 16. Since the abstract and conclusion claim robustness and absence of manual tuning, the authors should report at least a small sensitivity study over these parameters and include noisy simulations; otherwise the robustness claim is not yet established.","section":"Section 3.1 and Section 2.3"}],"minor_comments":[{"comment":"The section heading contains a typo, 'Conculsion', which should be 'Conclusion'.","section":"Section 4"},{"comment":"There are several typographical errors, including 'Institude' in the first affiliation, 'caputured' in Section 2.1, and 'verctors' in Section 2.1; these should be corrected.","section":"Throughout"},{"comment":"The summation in Eq. (4) runs over k = 1 to N, but N was previously used for the total number of object pixels and K for the number of scan positions; please use K consistently for scan positions.","section":"Section 2.1, Eq. (4)"},{"comment":"The function f is not defined explicitly; the text says it 'represents the incorporation of optimization variables into the model', but it should be stated precisely how xi and phi enter relative to the operators T and C, so that the model is reproducible.","section":"Algorithm 1, line 5 and Section 2.3"},{"comment":"The abstract's phrase 'average deviation below 0.5 pixels' should be stated as a mean over scan positions and complemented with a maximum or percentile, since for a = 5 Section 3.1 reports that most estimates remain within 1 pixel while only the mean is below 0.5 pixels.","section":"Abstract and Section 3.1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does one concrete thing: it puts the diffraction-pattern crop shift into the set of learnable parameters in automatic-differentiation ptychography. That specific variable is not in the cited AD frameworks, which already optimize scan positions and tilt angles. So the novelty is real but modest — a natural extension rather than a new paradigm. The motivation is well laid out, especially Figure 2, which shows how shifts distort the inverse TPC and broadband monochromatization. The implementation is straightforward, and the EUV experiment, with three discrete harmonics and a reflective geometry, is a nontrivial test bed.\n\nThat said, the central quantitative claim needs a closer look. The <0.5-pixel average deviation is demonstrated only in simulation, and the stress-test concern is on point: the ground-truth offsets appear to be generated with the same differentiable shift operator f that Algorithm 1 optimizes. The text never states the interpolation or resampling used for the ground truth, but if it is the same function, the optimizer is inverting the exact model that produced the data. That is an inverse-crime validation. Real detector misalignment includes pixel-area integration, detector blur, and geometric distortion, which are not guaranteed to live in that translation family. Under model mismatch, the recovered xi could minimize the loss without equaling the true crop offset, and the 0.5-pixel bound has no demonstrated transfer.\n\nThe experimental section does not fix this. It shows improved fringe contrast and peak separation with shift correction, but there is no ground truth, so the improvement could come partly from the extra degrees of freedom or from joint optimization of other variables. The baseline is also weak: it compares against no correction, not against a standard centering method like cross-correlation or center-of-mass. Those would be a more meaningful benchmark.\n\nThese are real soft spots, but they are not fatal to the idea. The framework is sensible, the math is fine, and the paper does not overclaim in the body — the abstract's 'subpixel correction' is the only place where the inverse-crime simulation is presented as enough. A serious referee should ask for a mismatched-model simulation (e.g., generate data with one interpolation kernel, optimize with another), a comparison to centering baselines, and ideally a release of code and data. The paper deserves peer review and is likely citable after those additions. I would not desk-reject it; I would send it out and request revisions.","headline":"A useful but incremental AD-ptychography extension whose subpixel-accuracy claim currently rests on a matched-model simulation; the real-data test is only qualitative.","tokens_in":12303,"tokens_out":2022,"would_cite":true,"duration_ms":26481,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["42.30.Rx"],"model":"deepseek-v4-flash","headline":"This paper claims that cropping-shift misalignment in ptychography can be corrected to subpixel accuracy by making the shift an optimizable parameter in an automatic-differentiation reconstruction loop.","keywords":["ptychography","automatic differentiation","diffraction pattern alignment","subpixel correction","reflection ptychography","broadband ptychography","EUV imaging","phase retrieval"],"falsifier":"Take a recorded diffraction pattern, crop it with known sub-pixel offsets generated by a shift model that includes a different interpolation kernel or pixel-averaging than $f(\\cdot,\\xi)$, run the proposed AD reconstruction, and compare the recovered offsets to the ground truth: if the mean deviation exceeds 0.5 pixels, the claimed accuracy is specific to the model rather than to real misalignment.","tokens_in":11320,"feed_emoji":"🔬","tokens_out":3332,"duration_ms":38602,"temperature":0.7,"pith_summary":"This paper argues that the misalignment between the crop window and the zero-order peak of a ptychographic diffraction pattern can be fixed automatically, inside the reconstruction itself, by treating the shift as a learnable parameter in an automatic-differentiation descent. If correct, it removes a manual preprocessing step that currently degrades reflection and broadband ptychography, where even small shifts cause structural distortion rather than simple translation. Simulations with preset offsets up to 5 pixels recover the true shifts with mean deviation below 0.5 pixels, and EUV experiments show sharper fringes and better contrast after correction. The intended payoff is a ptychographic workflow that jointly refines object, probe, and pattern alignment without heuristic or manual tuning.","feed_headline":"Crop misalignment in ptychography fixed to under half a pixel","feed_subtitle":"Automatic differentiation turns the cropping offset into a learnable parameter, sharpening EUV reconstructions.","key_machinery":"The load-bearing object is the differentiable shift operator $f(\\cdot,\\xi)$ applied to the modelled diffraction intensity in the algorithm's main loop, inserted between the forward propagation (together with optional tilted-plane transform $T$ and chromatic scaling $C$) and the L2 loss with TV regularization. Because the whole pipeline is expressed with automatic differentiation, the gradient of the loss with respect to the shift parameters flows back and updates them with Adam, making the cropping offset a refined quantity rather than a fixed preprocessing choice.","core_discovery":"The central claim is that subpixel diffraction-pattern shifts can be corrected by backpropagation while reconstructing the object and probe, by including the cropping shift vector $\\xi$ as an ordinary optimization variable in the AD-based loss. The paper demonstrates on three forward models—reflection, broadband, and combined reflection-broadband—that this recovers per-pattern offsets up to 5 pixels with an average deviation below 0.5 pixels, and that correcting the shift simultaneously improves reconstruction quality. In EUV experiments, enabling the shift correction sharpens reconstructed fringes and increases intensity contrast compared with reconstructions from coarsely cropped patterns, with further improvement when all parameters are jointly optimized.","pith_inferences":["A natural test of the method's limits is to compare recovered shifts against ground truth generated by a physically distinct shift model—for example detector pixelation and interpolation kernels different from the operator used in reconstruction—to see whether the sub-0.5-pixel accuracy persists.","The per-pattern shift parameter could be regularized jointly across scan positions, which may help when individual diffraction patterns are noisy or have low signal.","The reported accuracy suggests the approach could also serve as a diagnostic for residual systematic alignment errors in already-collected ptychography data, without reacquiring the measurements."],"forward_implications":["Users can skip manual centering of diffraction patterns because the reconstruction itself absorbs the alignment error.","Reflection and broadband configurations, where misalignment causes nonlinear distortion rather than a simple translation, stand to gain most from the correction.","The same AD loop can simultaneously optimize other hardware parameters such as scan positions, tilt angles, and background, as demonstrated in the EUV experiment.","The method is modular: changing the forward model only requires recompiling the differentiable graph, not hand-derived update rules.","Recovering subpixel offsets during reconstruction also provides a quantitative check on the preprocessing stage of existing ptychography datasets."],"supporting_citations":[{"why":"Provides the AD-based reflection ptychography framework whose loss structure and probe-object optimization this paper extends.","marker":"[4]"},{"why":"Supplies the tilt-angle optimization scheme and reflection geometry that motivate integrating pattern offsets as learnable variables.","marker":"[7]"},{"why":"Establishes the broadband coherent diffractive imaging model whose chromatic scaling operator C is reused here.","marker":"[28]"},{"why":"Documents the information loss and modeling error in numerical monochromatization that limit broadband reconstruction accuracy.","marker":"[29]"},{"why":"Introduces automatic differentiation ptychography as a general optimization framework for phase retrieval.","marker":"[32]"},{"why":"Defines the Adam optimizer used to update object, probe, and shift parameters.","marker":"[37]"},{"why":"Provides the automatic differentiation engine used to compute exact gradients through the forward model.","marker":"[38]"}],"fun_headline_variants":["Autodiff learns subpixel crop shifts in ptychography","Backprop corrects diffraction-pattern shifts to <0.5 px","Ptychography: learnable shift beats crop misalignment","Subpixel shift correction in ptychography via autodiff","AD-based crop shift refinement sharpens EUV ptychography"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The simulations create the misalignments with the same differentiable shift operation the reconstruction uses, so the validation does not test whether real-world shifts, which involve pixel response and interpolation, follow that exact translation model.","fun_headline_variants_meta":{"raw":{"variants":["Autodiff learns subpixel crop shifts in ptychography","Backprop corrects diffraction-pattern shifts to <0.5 px","Ptychography: learnable shift beats crop misalignment","Subpixel shift correction in ptychography via autodiff","AD-based crop shift refinement sharpens EUV ptychography"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1524,"prompt_tokens":918,"completion_tokens":606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":517}},"tokens_in":534,"tokens_out":606,"duration_ms":6745,"temperature":1.0,"reasoning_tokens":517,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:05:54.402576+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a recorded diffraction pattern, crop it with known sub-pixel offsets generated by a shift model that includes a different interpolation kernel or pixel-averaging than $f(\\cdot,\\xi)$, run the proposed AD reconstruction, and compare the recovered offsets to the ground truth: if the mean deviation exceeds 0.5 pixels, the claimed accuracy is specific to the model rather than to real misalignment.","supporting_citations":[{"cited_title":"Wavelength-multiplexed multi-mode euv reflection ptychography based on automatic differentiation,","cited_arxiv_id":null,"evidence_quote":"Provides the AD-based reflection ptychography framework whose loss structure and probe-object optimization this paper extends."},{"cited_title":"Mitigating tilt-induced artifacts in reflection ptychography via optimization of the tilt angles,","cited_arxiv_id":null,"evidence_quote":"Supplies the tilt-angle optimization scheme and reflection geometry that motivate integrating pattern offsets as learnable variables."},{"cited_title":"Broadband coherent diffractive imaging,","cited_arxiv_id":null,"evidence_quote":"Establishes the broadband coherent diffractive imaging model whose chromatic scaling operator C is reused here."},{"cited_title":"Broadband ptychographic imaging with an accurately sampled spectrum,","cited_arxiv_id":null,"evidence_quote":"Documents the information loss and modeling error in numerical monochromatization that limit broadband reconstruction accuracy."},{"cited_title":"Adp: Automatic differentiation ptychography,","cited_arxiv_id":null,"evidence_quote":"Introduces automatic differentiation ptychography as a general optimization framework for phase retrieval."}],"review_version":1}