{"id":"8eb63c21-80d8-4fc2-8861-0c3deee93d54","arxiv_id":"2608.09237","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A data-driven sensor placement framework reconstructs simulated AtLAST surface deformations to below 2.7 µm rms with 50 sensors, outperforming a Zernike basis.","lead":"This paper presents a method for choosing where to place deformation sensors on a large radio telescope's backup structure, using computer models of how the dish bends. It reports that 50 sensors can reconstruct the simulated surface to within a few micrometers, which would make an active surface for a 50-meter submillimeter telescope more feasible.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LOOCV leaks the held-out load case into sensor placement, so the 2.7 µm worst case may be optimistic until a nested placement CV is run.","rationale":"The paper is clear and internally consistent, and the POD-versus-Zernike comparison is fair for the stated task. The paper explicitly notes its own limitations: single noise realizations and the absence of inertial and slewing loads from the load-case library. The most load-bearing weakness is not external representativeness but the LOOCV leakage, because it directly affects the headline 2.7 µm number under the paper's own assumptions. The reader's weakest assumption (load-case representativeness) is real but is a known limitation of any data-driven basis; the sensor-selection leak is a correctness issue within the stated experiment. A strict nested CV is cheap and decisive. If it passes, the central claim is credible for design guidance; if it fails, the 30 to 50 sensor recommendation is unjustified. No verdict change beyond the reader's conditional is needed; the condition should be made explicit and tested.","tokens_in":6677,"tokens_out":5598,"duration_ms":61155,"concrete_test":"Re-run the N=50, k=20, σ=5 µm experiment with a fully nested LOOCV: for each of the 100 folds, build the POD basis from the 99 training snapshots only, run greedy I-optimal sensor selection with swap refinement on that training basis, then reconstruct the held-out case from its simulated readings. Record median and worst-case rms across folds. If the worst case remains near 2.7 µm and below 5 µm, the leak is negligible and the central claim stands; if it rises by more than about 1 µm or crosses the 5 µm target, the headline claim must be downgraded to 'conditional on sensor placement being optimized on the full deformation library.' For comparability, use the same noise realization as Fig. 4a, then repeat with at least 10 independent noise draws.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central validation result (Fig. 4a, Sec. 3.1) is produced by a leave-one-out scheme in which, per Sec. 2.5, only the POD basis is recomputed on the M-1 training snapshots while 'sensor positions determined once on the full set are retained.' The greedy I-optimal sensor selection therefore used the held-out load case when choosing where to measure. This is not a violation of the paper's stated assumptions, but it is a selection leak: the sensor layout is informed by the very deformation pattern being reconstructed. The 2.7 µm worst case is thus an estimate for deformations whose sensor layout was optimized with knowledge of them, not for a genuinely unseen deformation. Since each hold-out is only 1/100 of the library, the magnitude may be small, but the paper provides no bound; moreover, the statement that the 50-sensor configuration 'reconstructs all load cases' depends on this optimistic protocol. A strict nested LOOCV—recomputing both the POD basis and the greedy sensor selection inside each fold—would settle whether the reported number survives.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a data-driven framework for placing displacement sensors on the primary reflector of the AtLAST radio telescope and for reconstructing the full deformation field from the sparse sensor readings. A truncated Proper Orthogonal Decomposition (POD) basis is extracted from 100 finite-element load cases (gravity, thermal, wind); sensor locations are chosen greedily by I-optimality with subsequent swap refinement; and the reconstruction is performed by least-squares fitting of the noisy simulated sensor readings to the basis. The central numerical claim is that with 50 sensors, 20 POD modes, and 5 μm rms sensor noise, leave-one-out cross-validation yields per-load-case reconstruction residuals below 2.7 μm rms (median 2.4 μm), meeting the stated 5 μm BUS-level target, while an annular Zernike basis fails on the large gravitational deformation cases. The paper is clearly written and explicitly acknowledges several limitations, including the fact that the evaluation concerns the Back-Up Structure only and that inertial and slewing loads are not yet included in the load-case library.","tokens_in":6938,"tokens_out":6818,"duration_ms":69094,"significance":"If the quantitative claim survives a stricter validation protocol, the result is practically significant for the AtLAST active-surface concept: it suggests that a modest number of displacement sensors, on the order of 30 to 50, may suffice for closed-loop surface reconstruction, and it quantifies the advantage of a data-driven POD basis over a conventional analytic Zernike basis. The manuscript has clear strengths: the comparison with an analytic basis is illuminating, the authors are transparent about the BUS-only scope, and the figures are consistent with the reported numbers. The paper does not ship code or data, so independent reproduction is not immediate, but the methodological description is sufficiently detailed to be implementable. The main open question is whether the headline residual is robust under a fully nested cross-validation and under repeated noise realizations.","major_comments":[{"comment":"The leave-one-out protocol is not fully nested: while the POD basis is recomputed on the M−1 training snapshots, the text states that 'the sensor positions determined once on the full set are retained.' Consequently, the held-out load case has already participated in the greedy I-optimal sensor selection and swap refinement of Sec. 2.2. The headline worst-case residual of 2.7 μm is therefore an estimate for deformations whose sensor layout was informed by the very deformation pattern being reconstructed. This is a selection leak, not a violation of the stated assumptions, and its magnitude is not bounded in the paper. Please run a nested LOOCV in which both the POD basis and the sensor positions are recomputed inside each fold, or at least for a representative subset of folds, and report the resulting worst-case residual. The sentence in Sec. 2.5 claiming that this 'tests reconstruction on genuinely unseen deformations' is misleading in this respect.","section":"Sec. 2.5, Fig. 4a"},{"comment":"All quantitative results are based on a single Gaussian noise realization per load case. The reported worst-case residual of 2.7 μm is a single draw from the noise distribution, and the per-load-case reconstruction noise is not negligible when N=50 and σ=5 μm. The outlook acknowledges that repeated realizations are planned, but the central quantitative claim should not rest on one seed. Please repeat the evaluation over many independent noise realizations and report the distribution, or at least the worst-case residual over realizations, for the configurations shown in Figs. 4 and 5.","section":"Sec. 2.3 and Sec. 3, Figs. 4–7"},{"comment":"The 100 FEA load cases define the deformation space, and the reported residuals are meaningful only for deformations inside or near the span of that library. The paper lists inertial and slewing loads as future work, so the abstract's phrase 'reconstructs all load cases' refers to the simulated library, not to a guaranteed operational envelope. This conditionality should be stated more prominently, and, if feasible, a validation case from a distinct physical load category should be added to demonstrate that the POD subspace generalizes beyond the training library. Without such a test, the 2.7 μm number is best interpreted as an interpolation bound for the given load-case library rather than a prediction bound for arbitrary operational deformations.","section":"Sec. 2.1 and Sec. 4"}],"minor_comments":[{"comment":"The text states that the POD basis is evaluated for k ∈ {10,20,30}, but the legend of Fig. 5a shows k = 10, 20, and 50. Please reconcile the text and the figure.","section":"Sec. 3, Fig. 5"},{"comment":"The sentence 'As analysis showed, that results do not differ too much in terms of the optimization criterium' contains a grammatical error and a typo ('criterium' → 'criterion').","section":"Sec. 2.2"},{"comment":"The phrase 'This tests reconstruction on genuinely unseen deformations' should be qualified, as discussed in Major Comment 1; at minimum, the sentence should state that the sensor geometry is kept fixed and was selected using the full set.","section":"Sec. 2.5"},{"comment":"The abstract and introduction refer to 'surface deformation' of the primary reflector, while the evaluation is performed on the BUS nodes only. The paper does state this in Sec. 2.1, but the title and abstract could be slightly more explicit that the 2.7 μm figure concerns the BUS contribution, not the full optical surface error including panels and M2.","section":"Sec. 1 and Sec. 2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid engineering-methods contribution that fits the scope of an instrumentation or applied-optics journal. The main issue is the incomplete cross-validation protocol, which directly affects the headline number; this is fixable with a nested LOOCV run. The single-noise-realization concern is also fixable. I do not see a fundamental flaw that would require rejection, but the reported quantitative claim should not be accepted until the validation is strengthened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a clean, honest engineering methods paper. It takes POD-based sparse sensor placement (Manohar et al.) and I-optimal greedy selection (Fedorov), applies them to AtLAST's FEA load cases, and shows that 30-50 sensors with 20 POD modes can reconstruct the BUS deformation to below 2.7 um rms (worst case) under assumed 5 um sensor noise. The new contribution is the application to a radio telescope active surface and the explicit comparison showing a data-driven basis beats annular Zernikes on gravitational deformations. That comparison is credible and useful.\n\nThe paper is well structured, the math is standard but correctly applied, and it is upfront about limitations: BUS-only evaluation, single noise realization per load case, and the need to add inertial/slewing loads later. The figures match the claims.\n\nWhere it is soft: the LOOCV protocol leaks the held-out load case into sensor placement. Per Sec. 2.5, the sensor positions are determined once on the full set and retained in each fold; only the POD basis is recomputed. So the reported worst-case 2.7 um is for layouts that have seen the held-out case in the selection step. The stress-test note is right: a nested LOOCV that reselects sensors inside each fold would give an honest number. I would not be surprised if the result survives qualitatively--since each hold-out is only 1/100 of the library--but the paper does not currently show that. The single-noise-realization issue is acknowledged but still leaves the residual distributions without error bars. And no code or data is released, which makes the specific numbers hard to check independently. Representativeness of the 100 load cases for real operational conditions remains an open question, though the paper says the right things about it.\n\nDo I recommend engaging with this paper? Yes. It deserves peer review. The method is established, but the application is timely for AtLAST and the comparison with Zernikes is worth publishing. A referee should ask for the nested CV and preferably code/data release, but the core pipeline and headline result look plausible.\n\nHappy to discuss over coffee.\n\nBest.","headline":"Clean engineering application of sparse-sensing methods to AtLAST; the headline worst-case number needs a nested LOOCV before I'd quote it, but the paper deserves refereeing.","tokens_in":7391,"tokens_out":3066,"would_cite":true,"duration_ms":26596,"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 establishes that a data-driven POD basis learned from 100 finite-element load cases, combined with greedy I-optimal sensor placement, reconstructs every load case of AtLAST's back-up structure to below 2.7 µm rms using 50…","keywords":["active surface","sensor placement","proper orthogonal decomposition","sparse sensing","surface reconstruction","AtLAST","radio telescope","finite element analysis"],"falsifier":"Add the inertial and slewing loads the paper names as future work to the finite-element library, rerun the leave-one-out pipeline, and check whether the held-out worst-case residual stays below 5 µm; alternatively, compare the POD reconstruction against photogrammetry or holography measurements on the real structure under varied elevation and wind, and look for any measured deformation mode with negligible overlap with the POD subspace.","tokens_in":6520,"feed_emoji":"📡","tokens_out":7581,"duration_ms":76370,"temperature":0.7,"pith_summary":"This paper asks how many deformation sensors an active radio-telescope surface really needs, and where they should go, if the goal is to reconstruct the full reflector shape in real time for closed-loop control. The answer it defends is that, for the 50 m AtLAST concept, a data-driven Proper Orthogonal Decomposition (POD) basis learned from 100 finite-element load cases lets 30–50 sensors reconstruct the back-up structure's deformation to below 5 µm rms, the paper's reconstruction target. With 50 sensors and 20 modes, every load case is reconstructed to a worst-case 2.7 µm rms under 5 µm sensor noise. This matters because the ≈20 µm surface-accuracy requirement at submillimeter wavelengths is unattainable passively, and direct full-surface metrology is impractical, so sparse reconstruction is the enabling step for active surface control.","feed_headline":"Fifty sensors reconstruct a 50 m telescope dish to 2.7 µm","feed_subtitle":"Data-driven POD placement beats Zernike modes, making active surface control feasible for AtLAST's 20 µm accuracy goal.","key_machinery":"The load-bearing machinery is the POD basis $\\mathbf{U}_k$: the leading $k$ left singular vectors of the mean-centered snapshot matrix $\\mathbf{S}_{\\mathrm{FEA}} \\in \\mathbb{R}^{P \\times M}$, which define the low-dimensional subspace in which all reconstructed deformations live. Sensor selection then operates on the information matrix $\\Psi_N^\\top \\Psi_N$ formed from the $N$ selected rows of the basis, using greedy I-optimality (aperture-averaged prediction variance) with pairwise-swap refinement. Reconstruction is ordinary least squares, $\\hat{\\mathbf{a}}_j = (\\Psi_N^\\top \\Psi_N)^{-1} \\Psi_N^\\top \\mathbf{d}_j$, followed by $\\hat{\\mathbf{s}}_j = \\Psi \\hat{\\mathbf{a}}_j$; leave-one-out cross-validation recomputes the POD basis from $M-1$ snapshots to test genuinely unseen load cases. The essential result is that at $N=50$, $k=20$, this inverse problem is well-conditioned for the POD basis and not for the annular Zernike basis.","core_discovery":"The central claim is that a structure-specific, data-driven POD basis dramatically outperforms the conventional analytic annular Zernike basis for sparse surface reconstruction of a large active reflector. With 50 sensors and 20 modes, the POD basis reconstructs all 100 load cases under leave-one-out cross-validation to a median of 2.4 µm and a worst case of 2.7 µm rms, whereas the Zernike basis reaches a comparable median of 3.7 µm but a worst case of 176.4 µm, driven entirely by gravitational load cases. The paper also finds that the 5 µm reconstruction target is first met at 30 sensors with 20 modes, and that even when gravitational cases are excluded from the statistics, no Zernike configuration up to 250 sensors and 50 modes meets the target for all load cases.","pith_inferences":["Going beyond the paper, the same POD-plus-greedy pipeline should transfer to other large structures, such as panel-level metrology on other telescopes, provided the load-case library remains representative of real operating conditions.","The paper's evaluation uses one noise realization per load case, so the reported 2.7 µm worst case may not be a stable bound; a natural extension is to run many noise draws and report confidence intervals, which the authors list as future work.","If gravity is as repeatable and elevation-dependent as the FEA suggests, the reconstruction approach may make multi-elevation holographic calibration unnecessary, but that claim needs verification on real measured deformation fields rather than simulated ones."],"forward_implications":["A 30–50 sensor active-surface system could satisfy the reconstruction requirement for AtLAST's back-up structure, replacing a hypothetical system with several hundred sensors.","The data-driven POD basis is necessary for this performance; annular Zernike modes cannot represent gravitational deformation shapes even with up to 250 sensors and 50 modes.","Reconstruction residuals fall as sensor count grows until the coefficient estimate becomes noise-limited, and the mode truncation $k$ sets an irreducible floor.","If gravitational deformation is captured by the sensors, holographic calibration of the surface may be needed at a single elevation angle rather than across the full elevation range.","The framework applies to any structure for which representative load cases can be computed, including measured photogrammetry or holography data as an alternative to finite-element analysis."],"supporting_citations":[{"why":"Defines the AtLAST concept and the ≈20 µm half-wavefront-error requirement that motivates active surface reconstruction.","marker":"[1]"},{"why":"Supplies the finite-element structural deformations and technical requirements flow-down that define the load-case library and the BUS grid.","marker":"[5]"},{"why":"Defines annular Zernike polynomials, the conventional analytic basis against which the POD basis is compared.","marker":"[6]"},{"why":"Provides the Noll single-index convention used to truncate and order the Zernike basis.","marker":"[7]"},{"why":"Establishes Gram-Schmidt-style orthonormalization on arbitrary apertures, used to restore orthogonality of the Zernike basis on the annular domain.","marker":"[8]"},{"why":"Provides the data-driven greedy sparse-sensor-placement method for reconstruction that the sensor selection builds on.","marker":"[9]"},{"why":"Supplies the theory of optimal experiments underlying the D/I/G-optimality criteria and the swap refinement step.","marker":"[10]"}],"fun_headline_variants":["POD sensors beat Zernike: 50 sensors hit 2.7 µm for AtLAST","50 sensors, 20 modes: AtLAST surface control within reach","POD placement slashes telescope surface error to 2.7 µm","Why Zernike fails: POD-based sensor layout wins for radio dishes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 100 simulated load cases (6 gravity, 16 thermal, 78 wind) are treated as standing in for every deformation the telescope will actually experience; if real-world loads produce shapes outside that library, the reported error bound no longer applies.","fun_headline_variants_meta":{"raw":{"variants":["POD sensors beat Zernike: 50 sensors hit 2.7 µm for AtLAST","50 sensors, 20 modes: AtLAST surface control within reach","POD placement slashes telescope surface error to 2.7 µm","Why Zernike fails: POD-based sensor layout wins for radio dishes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1376,"prompt_tokens":1014,"completion_tokens":362,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":278}},"tokens_in":630,"tokens_out":362,"duration_ms":5262,"temperature":1.0,"reasoning_tokens":278,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:57:37.279480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Add the inertial and slewing loads the paper names as future work to the finite-element library, rerun the leave-one-out pipeline, and check whether the held-out worst-case residual stays below 5 µm; alternatively, compare the POD reconstruction against photogrammetry or holography measurements on the real structure under varied elevation and wind, and look for any measured deformation mode with negligible overlap with the POD subspace.","supporting_citations":[{"cited_title":"The conceptual design of the 50-meter Atacama Large Aperture Submillimeter Telescope (AtLAST)","cited_arxiv_id":"2402.18645","evidence_quote":"Defines the AtLAST concept and the ≈20 µm half-wavefront-error requirement that motivates active surface reconstruction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the finite-element structural deformations and technical requirements flow-down that define the load-case library and the BUS grid."},{"cited_title":"Greve and M","cited_arxiv_id":null,"evidence_quote":"Defines annular Zernike polynomials, the conventional analytic basis against which the POD basis is compared."},{"cited_title":", title =","cited_arxiv_id":null,"evidence_quote":"Provides the Noll single-index convention used to truncate and order the Zernike basis."},{"cited_title":"Manohar and B","cited_arxiv_id":null,"evidence_quote":"Establishes Gram-Schmidt-style orthonormalization on arbitrary apertures, used to restore orthogonality of the Zernike basis on the annular domain."},{"cited_title":"Pukelsheim , title =","cited_arxiv_id":null,"evidence_quote":"Provides the data-driven greedy sparse-sensor-placement method for reconstruction that the sensor selection builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the theory of optimal experiments underlying the D/I/G-optimality criteria and the swap refinement step."}],"review_version":1}