{"id":"55555b47-ef67-443a-a5cd-2b8caf6602d6","arxiv_id":"2508.04648","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An AO deformable mirror can be trained to produce optimized sub-pixel phase shifts that, combined by a jointly-optimized network, super-resolve undersampled telescope images without hardware changes.","lead":"A computational imaging method uses a telescope's existing adaptive-optics deformable mirror to inject tiny learned sub-pixel shifts between exposures, then fuses the frames with a trained super-resolution network to recover detail from undersampled cameras. No new hardware is needed; simulations and an optical-bench replica show resolution and SNR gains over classic dithering baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'AO-unaware' assumption and Eq. (5) forward model are unverified; optimized high-order DM flats could perturb the loop and science PSF, undermining the central 'SR for free' claim.","rationale":"The reader's CONDITIONAL verdict is well-aligned with my read. The strongest load-bearing vulnerability is indeed the set of assumptions around the small phase shift and the Eq. (5) image formation model. The paper asserts that AO science is unaffected and that the induced phase can be modeled as a convolution of a static AO PSF with the learned phase, but it provides no quantitative evidence for either claim. Figure 8 shows that optimized phases are not just small tip/tilt offsets; they include significant high-order Zernike power, and the power increases with SR factor. Whether such flats perturb the AO loop or the science PSF is an empirical question that the paper leaves unanswered. The OOMAO evaluation with optimized flats loaded as the DM reference is a step in the right direction, but the paper does not report the resulting Strehl or the comparison between the actual simulated PSF and the Eq. (5) prediction. The bench experiment similarly lacks an unperturbed control and includes test-time adaptation on the same data, so it does not isolate the effect of the learned phases. These are addressable concerns rather than demonstrated errors, so the appropriate verdict remains CONDITIONAL rather than ACCEPT or REJECT. I partially agree with the reader because they identified both the small-phase-shift assumption and the Eq. (5) model as load-bearing; my emphasis integrates them as one coupled risk: if the optimized flats are not truly small, the training model is wrong and the claimed gains may not transfer.","tokens_in":14848,"tokens_out":8111,"duration_ms":108471,"concrete_test":"Using the same OOMAO configuration as Sec. 5.1, generate (A) nominal-flat and (B) optimized-flat time-averaged science PSFs exactly as described in Sec. 5, and also compute the Eq. (5) prediction for the same P_i and S_i. Compare (B) to (A) and to the prediction, reporting Strehl ratio, PSF FWHM, and normalized L2 error. If the optimized-flat PSF differs from the nominal PSF by more than a few percent in Strehl/FWHM, or if the Eq. (5) prediction error is larger than the simulation noise floor, the 'AO-unaware' assumption and the training forward model are not supported, and the claimed SR gains need reassessment. A corresponding bench check would log WFS slopes and record a pinhole PSF with and without the optimized flats.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central benefit—super-resolution with no impact on AO/science—rests on Sec. 4.1's assertion that the optimized phase offsets leave the AO loop 'entirely as usual.' This is never measured. The learned flats in Fig. 8 contain substantial modal power in coma, trefoil, spherical, and other non-tip/tilt modes, with power rising with SR factor; these are not pure sub-pixel shifts. Loading such flats changes the WFS reference: the loop servos to the new setpoint, and any offset component outside the WFS/DM spatial bandwidth is not maintained, while the in-band component appears in the science path as added aberration. No WFS residual, loop gain, or Strehl/rms wavefront error is reported for the offset flats, in either OOMAO or the bench setup. Footnote 1's 'if necessary, impose constraints' is a contingency, not evidence. In addition, the training forward model Eq. (5) assumes each sub-exposure PSF equals |P_i ⊛ e^{jS_i}|^2, i.e., a static AO PSF convolved with the induced phase. For a time-varying atmosphere, the incoherent average of |P(t) ⊛ h_S|^2 does not factor as |⟨P⟩ ⊛ h_S|^2 when S is non-negligible; cross-terms couple the residual phase and S. If the true PSF under the optimized flats differs materially from Eq. (5), the jointly optimized phases and network are trained on the wrong model, and Table 1/Fig. 4 gains may not transfer. The bench demo cannot rule this out: it uses modified TENT adaptation on the same data and reports no unperturbed-flat control or PSF/Strehl comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a computational imaging method for astronomical telescopes equipped with adaptive optics. The idea is to use the deformable mirror to apply learned phase diversity across N sub-exposures, producing sub-pixel information that is then combined by a trained multi-image super-resolution network (or a linear shift-and-add reconstructor). The phase profiles and the reconstructor are jointly optimized offline against a forward model built from AO PSFs. The method is evaluated in OOMAO simulations for SR factors 2x, 4x, and 8x and with N=2,4,6,8 sub-exposures, and on an optical bench with an ALPAO DM and Shack-Hartmann WFS, with MTF comparisons. The central claims are that this yields SNR gains up to 12 dB over non-AO baselines and that the AO loop and science data are essentially unaffected by the injected phases.","tokens_in":15226,"tokens_out":8178,"duration_ms":90116,"significance":"The core idea is novel and practically attractive: if the DM can inject diversity without degrading science, existing AO-fed instruments could obtain super-resolution with no hardware changes. The simulated evaluation is a genuine generalization test—phases are optimized against a distribution of training PSFs and then applied as the DM flat in a fresh OOMAO simulation (Sec. 5), so the reported gains are not fitted directly to the test images. The bench prototype with a real DM, WFS, and phase screen is a substantial experimental effort. However, the headline quantitative claim is not supported by the reported tables, and the central 'AO-unaware' assumption is asserted rather than demonstrated. Several internal inconsistencies and missing experimental controls prevent the paper from being accepted in its current form.","major_comments":[{"comment":"The abstract's headline 'significant SNR improvements of up to 12 dB' is not reproducible from the reported data. In Table 1, the largest gain over the stated CNN baseline ('No Learned Phase') is 43.13 - 33.72 = 9.41 dB at 2x. No 12 dB value appears in any table or figure. Please either correct the headline or identify the exact baseline and configuration that produce 12 dB.","section":"Abstract; Table 1"},{"comment":"The central claim that the AO loop runs 'entirely as usual' and that the science data are unaffected is not verified. The optimized flats in Fig. 8 contain substantial power in non-tip/tilt Zernike modes (coma, trefoil, spherical, etc.), and the absolute amplitude of these phases is never reported. No wavefront-sensor residuals, loop gain, Strehl ratio, or RMS wavefront error are given for the optimized flats in either the OOMAO simulations or the bench experiment. Eq. (5) models each sub-exposure PSF as a static AO PSF modified by the induced phase; for a time-varying atmosphere, the long-exposure PSF is the time average of |p(t) * h_S|^2, which does not factor as |<p> * h_S|^2 when S is non-negligible. Since the learned phases are evidently not negligible, the training forward model may be misspecified, which would undermine the optimized phases and the Table 1 gains. Please provide di","section":"Sec. 4.1; Eq. (5); Sec. 5.1; Fig. 8"},{"comment":"There is an inconsistency between the two parts of Table 1. With N=4, the 'Scale Factor' section reports CNN Tip/Tilt (Classic) as 39.56 dB at 2x and 33.46 dB at 4x; the 'Number of Exposures' section reports 33.89 dB at N=4 (and for Non-Modal it reports 34.24 dB at N=4, matching the 4x column, while Tip/Tilt Classic does not). The super-resolution factor for the 'Number of Exposures' section is not stated. This must be clarified and reconciled before the cross-method comparisons can be assessed.","section":"Table 1"},{"comment":"The bench experiment does not independently confirm the simulation claims. Only qualitative images and MTF curves are reported; no PSNR/SNR values are given for the hardware data. The network is adapted to the bench data with modified TENT using the same test data, and no control with unperturbed flats or with the baseline CNN is shown. Without a quantitative comparison against a no-learned-phase or classic-dithering baseline on the same hardware, the experimental section cannot substantiate the 'up to 12 dB' claim or the 'no impact on science' assertion.","section":"Sec. 6; Sec. 6.1; Supplemental Sec. 10"}],"minor_comments":[{"comment":"P_i is used inconsistently as both a PSF and a complex amplitude. Please define the complex amplitude p_i and the phase-kernel h_S explicitly, and write the convolution as p_i * h_S with |.|^2 for the PSF.","section":"Eq. (5)"},{"comment":"The text says 'Experimental results are shown in Sec. 5.1', but Sec. 5.1 is titled 'AO Simulation Environment'. The subsection numbering or cross-reference should be fixed.","section":"Sec. 5.1"},{"comment":"USFA should be USAF in the caption ('USFA resolution target').","section":"Fig. 6 caption"},{"comment":"Reference [58] contains a typo: 'adaptive ptics' should be 'adaptive optics'.","section":"References"},{"comment":"The modal calibration is described for a single mode, but the optimized flats are superpositions of many modes. Please validate that the full multi-mode flat is reproduced accurately on the bench.","section":"Supplemental Sec. 10"},{"comment":"The 'No Learned Phase' baseline is described in one place as 'simple bilinear interpolation' and in another as a CNN with unshifted PSFs. Please state precisely what the baseline in Table 1 is.","section":"Sec. 4.2 / Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong, interesting idea and the simulation protocol is a genuine generalization test, but the abstract overstates the reported gains and the central 'AO-unaware' premise is unverified. The Table 1 inconsistency and the lack of quantitative bench controls are fixable, but the forward-model concern in Eq. (5) requires either a theoretical justification with constrained phase amplitudes or new validation data. The manuscript fits the journal's scope and has no apparent novelty or attribution issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a genuine new idea—using the AO deformable mirror's flat to inject learned phase distortions for multi-image super-resolution, co-optimized with a CNN—and the simulation evidence is more careful than most. But the headline '12 dB' and the claim that science data are unaffected are not backed up, and the forward model's range of validity is unexamined.\n\nWhat is good: the joint optimization of DM phases and MISR network is new, the Zernike modal analysis (Fig. 8) is a nice diagnostic, and the bench-top AO replica is a real effort. The simulation evaluation is a proper generalization test: phases are trained on a distribution of PSFs, then applied as the DM flat in fresh OOMAO simulations, so the reported gains are not fitted to test images.\n\nSoft spots, in rough order of severity. (1) The abstract's 'up to 12 dB' cannot be reproduced from Table 1; the largest gain shown is 9.41 dB. (2) No error bars, so we don't know whether the differences between phase representations are significant, especially at 8x. (3) The bench demo uses cropped images and a modified TENT that updates batch-norm on the test data itself; there is no unperturbed-flat control and no PSF/Strehl comparison, so the 'AO runs entirely as usual' claim is measured nowhere. (4) The Eq. (5) forward model assumes the time-averaged PSF with an added phase is the convolution of the averaged AO PSF with that phase; for non-negligible S, cross-terms with the residual turbulence break that. The simulation and bench both inherit this assumption. (5) Minor but important for reproducibility: main text says 397-actuator DM, supplement says 17x17 (289) DM order; r0 and wind speeds are given in units that must be wrong (cm for r0, km/s for wind). These are fixable but shouldn't be left as is.\n\nThe central method is plausible, and the paper is worth taking seriously. It's a methods contribution for AO-fed instrumentation; a careful referee should ask for corrected numbers, error bars, a statement of the forward model's limits, and at least a measurement of the WFS residual or science PSF with the optimized flats applied. I'd send it to review.","headline":"Real idea and solid simulation evidence, but the 12 dB headline and the claim that the AO loop is unaffected need to be backed up before the paper is as good as it looks.","tokens_in":15814,"tokens_out":3636,"would_cite":true,"duration_ms":42708,"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":"The paper claims that an adaptive optics deformable mirror can inject learned phase offsets that create sub-pixel shifts across sub-exposures, letting a jointly optimized network super-resolve undersampled telescope images by 2x to 8x with","keywords":["super-resolution","adaptive optics","deformable mirror","computational imaging","undersampled imaging","multi-image super-resolution","wavefront phase optimization","astronomical instrumentation"],"falsifier":"On an operational AO telescope, command the optimized phase profiles as the mirror flats for N sub-exposures and concurrently record wavefront-sensor residuals, loop gain, and the long-exposure science PSF/Strehl ratio. If the residuals or Strehl change measurably relative to the nominal flat while the induced sub-pixel shifts are present, the Small Phase Shift premise is refuted. As a complementary check, measure the centroid of a point source after each learned flat is applied: each sub-exposure must carry a distinct, known sub-pixel displacement matching what the network was trained on.","tokens_in":14692,"feed_emoji":"🔭","tokens_out":11698,"duration_ms":126253,"temperature":0.7,"pith_summary":"Astronomical cameras that image a wide field or use large pixels undersample the telescope's diffraction-limited point spread function, throwing away resolution. This paper claims that the deformable mirror already inside an adaptive optics system can recover that lost resolution: by commanding the mirror to add small learned phase distortions during a burst of sub-exposures, each frame lands on the detector with a slightly different sub-pixel shift, and a jointly trained upsampling network fuses the burst into a super-resolved image. The authors test this in realistic AO simulations and on a bench-top telescope replica, reporting 2x, 4x, and 8x super-resolution with four sub-exposures, higher PSNR than unoptimized or classic tip/tilt dithering baselines, and improved modulation transfer function. If it holds on sky, existing observatories could gain resolution without new optics and without spending extra on-sky time.","feed_headline":"Existing AO mirrors can deliver up to 8x sharper telescope images","feed_subtitle":"Learned mirror phases add sub-pixel shifts to each exposure, letting AO cameras reconstruct up to 8x finer detail without new hardware.","key_machinery":"The engine is the joint optimization in Eq. (6): $\\min_{\\theta,S} \\| f_\\uparrow( f_\\downarrow( \\{I_i\\} | \\theta) ) - O \\|_1$, with sub-exposures generated by Eq. (5), $I_i = |O|^2 \\circledast |P_i \\circledast e^{jS_i}|^2 + \\eta$. Here $S_i$ is the trainable phase profile applied by the deformable mirror, $P_i$ is the AO-corrected PSF of the $i$-th sub-exposure, $f_\\uparrow$ is a multi-image upsampler (a residual CNN or a linear shift-and-add), $f_\\downarrow$ is the known downsampling, and $\\eta$ is noise. The mirror phase is the key object: it is not restricted to the planar tip/tilt of traditional dithering, and the optimizer is free to sculpt high-order wavefront modes that create sub-pixe","core_discovery":"The central claim is that the AO deformable mirror can act as a programmable dithering device for super-resolution. For each of $N$ sub-exposures, the mirror adds an optimized static phase $S_i$ to the wavefront; the image-formation model becomes $I_i = |O|^2 \\circledast |P_i \\circledast e^{jS_i}|^2 + \\eta$, where $P_i$ is the AO-corrected PSF. The phases and the upsampling network are trained together by minimizing the $\\ell^1$ error between the jointly upsampled reconstruction and the true high-resolution image (Eq. (6)). In simulation, directly optimized mirror shapes reach 43.13 dB PSNR at 2x with $N=4$, versus 33.72 dB for the same network with no learned phase; classic and optimized ti","pith_inferences":["Table 1 hints that with learned reconstruction the return on additional exposures keeps growing (37.48 dB at N=8 for directly optimized mirror shapes) while linear reconstruction plateaus near 26.7 dB regardless of N; a plausible inference is that learned reconstructors are the only ones that fully monetize added phase diversity.","A quantitative on-sky test of the 'Small Phase Shift' premise—recording wavefront-sensor residuals and science Strehl while cycling optimized flats—would settle whether the super-resolution truly comes for free; the paper defers this guarantee to a footnote about constraining phase magnitudes if needed.","The choice of phase parameterization carries a deployment trade-off implicit in the paper: directly optimized mirror shapes give the best simulation results but require precise influence-function calibration, whereas modal phase bases are easier to command on real systems.","Since the phase profiles are learned for a telescope's own PSF statistics, the same framework could plausibly be re-optimized for multi-conjugate AO systems by assigning different phase rates to different deformable mirrors, a direction the paper lists as future work."],"forward_implications":["Undersampled instruments on single-conjugate AO telescopes could observe as four or more sub-exposures, each with a different learned mirror flat, and reconstruct a 2x–8x finer image from the same total exposure time.","Because phase optimization is done offline from simulated or telemetry-derived PSFs, deployment requires no hardware changes and no added on-sky calibration.","The benefit does not depend on deep learning alone: a linear shift-and-add reconstructor also improves with optimized phases, so simpler pipelines can capture part of the gain.","The optimized phases contain substantial high-order wavefront content, not just tip/tilt, so the full shape space of the deformable mirror is the resource being exploited.","If the central claim is correct, an on-sky test on a science-class telescope should reproduce the bench-top MTF gains, since the bench-top replica already uses the same AO loop and reference-flat mechanism."],"supporting_citations":[{"why":"The authors' earlier proposal of AO-mediated sub-pixel super-resolution; this paper extends it from fixed shifts to end-to-end optimized phases.","marker":"[54]"},{"why":"Supplies the AO simulation environment used to generate realistic time-varying PSFs for training and testing.","marker":"[14]"},{"why":"Defines the classical dithering/shift-and-add reconstruction that serves as the linear baseline and comparison point for optimized phases.","marker":"[23]"},{"why":"Provides the residual-CNN architecture that is modified into the multi-exposure upsampling network.","marker":"[39]"},{"why":"One of the two image datasets used as high-resolution ground truth for training the phases and the network.","marker":"[1]"},{"why":"The second training dataset of real image pairs, used to make the upsampler robust to natural scenes.","marker":"[30]"}],"fun_headline_variants":["AO mirror dithering produces super-resolved telescope images","Learned mirror phases add sub-pixel shifts for sharper views","Existing AO hardware delivers super-resolution via learned dithering","Break the resolution-FoV tradeoff with adaptive-optics dithering","Telescope AO system re-purposed for 12 dB sharper imaging"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is the 'Small Phase Shift' assumption: the learned deformable-mirror phases are small enough that the AO control loop and the science exposure run exactly as usual, so super-resolution costs nothing in wavefront-correction quality or science signal; if the injected phases perturb wavefront sensing or blur the science PSF, the method's central benefit collapses.","fun_headline_variants_meta":{"raw":{"variants":["AO mirror dithering produces super-resolved telescope images","Learned mirror phases add sub-pixel shifts for sharper views","Existing AO hardware delivers super-resolution via learned dithering","Break the resolution-FoV tradeoff with adaptive-optics dithering","Telescope AO system re-purposed for 12 dB sharper imaging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000327,"raw_usage":{"total_tokens":1717,"prompt_tokens":844,"completion_tokens":873,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":783}},"tokens_in":588,"tokens_out":873,"duration_ms":10085,"temperature":1.0,"reasoning_tokens":783,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:51:38.817604+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On an operational AO telescope, command the optimized phase profiles as the mirror flats for N sub-exposures and concurrently record wavefront-sensor residuals, loop gain, and the long-exposure science PSF/Strehl ratio. If the residuals or Strehl change measurably relative to the nominal flat while the induced sub-pixel shifts are present, the Small Phase Shift premise is refuted. As a complementary check, measure the centroid of a point source after each learned flat is applied: each sub-exposure must carry a distinct, known sub-pixel displacement matching what the network was trained on.","supporting_citations":[{"cited_title":"Adaptive optics mediated sub-pixel super resolution","cited_arxiv_id":null,"evidence_quote":"The authors' earlier proposal of AO-mediated sub-pixel super-resolution; this paper extends it from fixed shifts to end-to-end optimized phases."},{"cited_title":"Object-oriented matlab adaptive optics toolbox","cited_arxiv_id":null,"evidence_quote":"Supplies the AO simulation environment used to generate realistic time-varying PSFs for training and testing."},{"cited_title":"Drizzle: A method for the linear reconstruction of undersampled images","cited_arxiv_id":null,"evidence_quote":"Defines the classical dithering/shift-and-add reconstruction that serves as the linear baseline and comparison point for optimized phases."},{"cited_title":"Enhanced deep residual networks for sin- gle image super-resolution","cited_arxiv_id":null,"evidence_quote":"Provides the residual-CNN architecture that is modified into the multi-exposure upsampling network."},{"cited_title":"Ntire 2017 challenge on single image super-resolution: Dataset and study","cited_arxiv_id":null,"evidence_quote":"One of the two image datasets used as high-resolution ground truth for training the phases and the network."},{"cited_title":"Imagepairs: Realistic super resolu- tion dataset via beam splitter camera rig","cited_arxiv_id":null,"evidence_quote":"The second training dataset of real image pairs, used to make the upsampler robust to natural scenes."}],"review_version":1}