{"id":"70106967-acd5-44e2-9625-9ac11dbb8798","arxiv_id":"2502.00101","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"The Teledyne COSMOS-66 CMOS sensor, after a pixel-by-pixel nonlinearity correction, reaches 2.9 e- effective read noise and 0.12 e-/pix/s dark current at -25 C, with QE peaking at 89%.","lead":"A team at MIT bench-tested and put on the sky a new large-format CMOS camera from Teledyne, measuring its noise, dark current, quantum efficiency, and pixel crosstalk. They found low-light nonlinearity, built a pixel-by-pixel correction, and demonstrated 2.9 electron read noise and accurate stellar photometry.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: X-ray gain assumption is cross-checked by PTC and on-sky photometry, so the central electron-unit claims hold up.","rationale":"The paper's central claim is an engineering characterization, and the strongest-claim read-noise and dark-current numbers are supported by a coherent chain: the nonlinearity is empirically calibrated, the calibration is shown to linearize fresh data, the effective noise scaling is predicted by the trap model and confirmed by PTC and on-sky noise curves, and the conversion gain is anchored by two independent methods. The X-ray pair-creation energy assumption is the only place where an external physical constant enters the absolute calibration; but because the PTC gain (a purely statistical measurement) agrees with the X-ray gain to ~1%, any error in that constant would have to be coincidentally mirrored by the PTC analysis. The on-sky absolute photometry adds a further consistency check. The residual concern is reproducibility: data and code are not public, so an independent team cannot re-run the analysis without requesting them. This justifies the reader's CONDITIONAL verdict but does not constitute a technical objection to the central claim. The stress-test found no reason to move the verdict.","tokens_in":19530,"tokens_out":11945,"duration_ms":128040,"concrete_test":"Request the representative dataset/code from the authors and recompute the X-ray conversion gain using the temperature-dependent electron-hole pair-creation energy for silicon at 248 K (e.g., from Lowe & Sareen 2007). Verify that the derived read noise remains 2.9 ± 0.2 e- and that the PTC gain and on-sky photometric slope shift by less than 2%. If the shift is larger, the headline electron-unit values need downward/upward revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most assumption-dependent step is the X-ray conversion gain (Sec. 4.2), which scales every electron-unit quantity. A bias here would shift read noise, dark current, QE, and full well together. However, the paper supplies two independent anchors that make this non-load-bearing: (1) the photon transfer gain, corrected for interpixel capacitance, agrees with the X-ray gain to within ~1% across all three modes (Table 3); (2) the on-sky photometric slope in Fig. 16a matches the independently known telescope effective area (0.14 m^2). The latter checks the product of QE and gain, and while a compensating error in the pair-creation energy and QE could in principle survive, the PTC agreement rules out a large gain error. The agreement between single- and multi-pixel X-ray event spectra after calibration (Sec. 4.2) also indicates complete charge collection in graded events. The remaining uncertainty from the assumed E/3.69 eV value (Lowe & Sareen, 2007) is a few percent at 248 K, far below the 0.2 e- uncertainty quoted for the 2.9 e- read noise. No internal inconsistency or unsupported empirical claim was found; the effective-vs-physical noise distinction is clearly labeled throughout.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Layden et al. present a laboratory and on-sky characterization of the Teledyne COSMOS-66, an 8120 x 8120 pixel, 10 μm back-illuminated CMOS sensor. The paper identifies a low-signal nonlinearity, develops and applies a per-pixel polynomial correction, and measures conversion gain via both Cd-109 X-ray spectra and photon-transfer curves. It reports effective read noise, dark current, quantum efficiency, nonuniformity, image lag, and crosstalk for three readout modes. The headline results are an effective read noise of 2.9 e− RMS in correlated multi-sampling mode at −25 °C, a dark current of 0.12 e−/pix/s, and quantum efficiency above 50% from 250 nm to 800 nm. An on-sky test at the WINTER 1 m telescope validates the photometric calibration and noise model.","tokens_in":19800,"tokens_out":15113,"duration_ms":144212,"significance":"This is a solid, well-cross-checked detector characterization. The two independent conversion-gain measurements (X-ray and PTC) agree to about 1% after the IPC correction, and the on-sky photometric slope matches the known telescope effective area, making the absolute electron-unit calibration credible. The authors clearly distinguish the physical read noise (0.77 e−) from the effective read noise (2.9 e−) introduced by the nonlinearity correction, which is important for correct interpretation. The trap model is explicitly presented as a plausible explanation rather than a proven mechanism, and the empirical calibration is validated by temporal stability and on-sky photometry. The measured performance and calibration methodology are relevant to astronomical instrument development.","major_comments":[],"minor_comments":[{"comment":"The text says 'the number of such pixels was very small for the CMS and HSHGGS modes, but for the HSHGGS mode, nearly the entire top 75 rows exhibited non-monotonic response'; this is internally inconsistent, and the first mode should likely be HSHGRS to agree with Table 4, which shows a 1.096% uncalibratable/lazy fraction for HSHGGS.","section":"Sec. 4.3"},{"comment":"The sentence 'As shown in Fig. 4c, this calibration worked equally well for the trap model data' appears to reference the wrong panel, since Fig. 4c shows the raw PTC while the corrected PTC is shown in Fig. 4d; please correct the reference.","section":"Sec. 3.1 (after Fig. 4)"},{"comment":"The phrase 'the the Large Synoptic Survey Telescope' contains a duplicated article and should be corrected.","section":"Sec. 1"},{"comment":"The frame-rate units appear as 'f ps' in Table 1's header and as 'f psmay' in the introduction; these should be typeset as 'fps'.","section":"Table 1 and Sec. 1"},{"comment":"The organization name is given as 'European Machine Vision Association (EMV A)' with a stray space; this should be 'EMVA'.","section":"Reference [27]"},{"comment":"The PTC conversion gains are explicitly corrected for the ~1.6% interpixel-capacitance effect, but the X-ray gains adopted for analysis are not; since IPC can in principle reduce the pulse height of single-pixel X-ray events, the authors should state whether the event-summing procedure already recovers the IPC-shared charge or whether a small systematic offset remains between the two gain scales.","section":"Secs. 4.2 and 4.8"},{"comment":"The dark-current model is described as an exponential with a doubling temperature of 5.2 °C plus an asymptotic value of 0.06 e−/pix/s, but the fit is not shown with uncertainties or a goodness-of-fit estimate; a brief figure or equation would clarify this result.","section":"Sec. 4.5"},{"comment":"The paper states that the data are not publicly available and only a representative subset is available on request; for reproducibility of the key calibration and noise results, making at least the calibration ramps and X-ray spectra publicly available would be helpful.","section":"Code, Data, and Materials Availability"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is within the journal's scope and the technical content is sound. The main limitations are the lack of public data and the reliance on an NDA-based root-cause explanation for the nonlinearity, but neither undermines the empirical results, which are cross-validated by independent methods. I recommend publication after minor revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first independent public characterization of the Teledyne COSMOS-66, and the central results hold up. The genuinely new bit is the low-signal nonlinearity (below 5% of saturation) and the pixel-by-pixel correction they built for it.\n\nWhat the paper does well: the gain measurements are cross-anchored. X-ray gain agrees with PTC gain to roughly 1% in all three modes after the IPC correction, and the on-sky photometric slope matches the effective 0.14 m^2 telescope area. The X-ray single-pixel versus multi-pixel spectra agree after calibration, which checks the charge-collection assumption. The noise model is validated on sky, and the conclusion is appropriately modest: with the current nonlinearity and crosstalk, the sensor does not beat large-format CCDs. That honesty follows from the data.\n\nSoft spots: the headline 2.9 e- read noise is an effective, post-calibration number; the raw physical noise is 0.7 e-. The authors are clear about this, but anyone citing the camera should not drop the qualifier. The trap model is two fitted parameters and is explicitly presented as plausible, not derived; it is not load-bearing. The calibration is a fit to the same illumination ramps used for some derived quantities, but the independent cross-checks make that a minor circularity, not a flaw. The bigger practical issue is that the data and code are not public—only available upon request. For a detector characterization paper that others will benchmark against, that limits independent verification. The cross-checks are strong enough that I do not think it sinks the paper, but it is a real limitation for a paper of this type.\n\nWho gets value: detector astronomers choosing sensors, anyone planning to use COSMOS, and instrument-builders who need to know about the trap-like nonlinearity. It deserves a serious referee. I would send it out, with the request that the authors post at least the representative dataset and code.","headline":"First independent characterization of the COSMOS-66: the low-signal nonlinearity and its calibration are genuinely new, and the cross-checked numbers hold up.","tokens_in":20369,"tokens_out":1704,"would_cite":true,"duration_ms":18013,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that a pixel-by-pixel nonlinearity calibration turns the COSMOS-66 CMOS sensor into a CCD-class astronomical imager, with $2.9\\,e^{-}$ read noise and $0.12\\,e^{-}/\\mathrm{pix/s}$ dark current at $-25\\,^{\\circ}\\mathrm{C}$.","keywords":["CMOS image sensors","detector characterization","large-format image sensors","read noise","nonlinearity correction","quantum efficiency","astronomical instrumentation","correlated multi-sampling"],"falsifier":"Compare the X-ray-derived conversion gain with a photodiode-calibrated photon flux at a known wavelength and recompute read noise and dark current; if the two gains disagree beyond the quoted errors, all electron-unit values shift in proportion.","tokens_in":19334,"feed_emoji":"🔭","tokens_out":8583,"duration_ms":75859,"temperature":0.7,"pith_summary":"This paper reports a complete laboratory and on-sky characterization of the COSMOS-66, a large-format CMOS camera with $8120 \\times 8120$ ten-micron pixels. The central result is that the sensor's pronounced low-light nonlinearity can be removed with a per-pixel polynomial calibration, after which the correlated multi-sampling mode achieves an effective read noise of $2.9\\,e^{-}$ and a dark current of $0.12\\,e^{-}/\\mathrm{pix/s}$ at $-25\\,^{\\circ}\\mathrm{C}$. The quantum efficiency exceeds 50% from 250 to 800 nm, peaking at $89\\%$ at 600 nm. On the sky, the calibrated camera produced stellar photometry whose noise matched the sum of effective read noise, effective shot noise, and scintillation, confirming CCD-comparable performance with faster readout and relaxed cooling. The paper also measures about 1.5-pixel FWHM optical crosstalk from photoelectron diffusion and proposes electron trapping as the mechanism behind the nonlinearity.","feed_headline":"Per-pixel fix brings 66-megapixel CMOS to 2.9 e− read noise","feed_subtitle":"After the linearity correction the COSMOS-66 posts 0.12 e−/pix/s dark current at −25°C and 89% peak QE.","key_machinery":"The load-bearing mechanism is the per-pixel nonlinearity calibration. For each pixel and operating mode, the camera's mean response to a stable, uniform 640 nm illumination ramp is fit by two polynomials—an 11th-order fit below 5% of saturation and a 5th-order fit above—and the fitted response is mapped onto the ideal linear response passing through the origin. Because the calibration is applied to all frames before computing noise, the sensor's noise properties are quoted as effective values: the correction amplifies low signal values, so the measured read noise and shot noise in the faint-signal regime are scaled up by about a factor of 4 and 2, respectively, compared to the raw values. A companion trap model reproduces the shape of the nonlinear response curve and the photon transfer curve, supporting the interpretation that low-signal electrons are lost to about 365 traps per pixel.","core_discovery":"The paper's central claim is that the COSMOS-66 CMOS sensor, once calibrated pixel by pixel, is a viable large-format astronomical imager whose performance is comparable to that of CCDs. The key finding is that the sensor's response is strongly nonlinear below roughly 5% of full well, stable over time, and well described by a model in which each pixel contains about $N_t = 365$ electron traps that capture passing electrons with probability $p = 0.002$. A per-pixel calibration, built from a 104-step illumination ramp and using an 11th-order polynomial below 5% saturation and a 5th-order polynomial above, linearizes the response across the dynamic range, lowers pixel response nonuniformity from $1.1\\%$ to $0.09\\%$, and makes the effective read noise $2.9\\,e^{-}$ RMS in the correlated multi-sampling mode while the shot noise at small signals is approximately $2\\sqrt{N_e}$. The reported absolute electron values (dark current $0.12\\,e^{-}/\\mathrm{pix/s}$ at $-25\\,^{\\circ}\\mathrm{C}$, peak quantum efficiency $89\\%$) follow from this calibration.","pith_inferences":["The trap model predicts a small bump in the photon transfer curve near $N_t \\approx 365$ electrons per pixel; measuring the PTC at several temperatures would test whether the trapping is thermally activated and let astronomers estimate trap energies from dark-current behavior.","The noise penalty pattern—effective read noise about four times physical and effective shot noise about twice the ideal at faint signals—implies that faint-object observations pay a 2× penalty until the traps are fixed, so exposure-time calculators for surveys should include this factor explicitly.","Because the crosstalk kernel is wavelength-dependent (blue light spreads more than red), a natural next step is to measure the full $5 \\times 5$ kernel and fit a diffusion model; that model could predict whether revised pixel designs with deeper depletion would reduce the spread before prototypes are built."],"forward_implications":["A calibrated COSMOS-66 can act as a drop-in large-format imager for surveys that want CCD-scale noise with higher frame rates and thermoelectric cooling instead of cryogenic cooling.","Photometric error budgets for this camera must use the effective read noise of $2.9\\,e^{-}$ and effective shot noise of about $2\\sqrt{N_e}$ at faint signals, not the raw $0.7\\,e^{-}$ specification.","The calibration absorbs most flat-fielding, reducing pixel response nonuniformity to $0.09\\%$, so data reduction can skip or simplify traditional flat-fielding steps.","If future pixel designs eliminate the nonlinearity and the ~1.5-pixel diffusion crosstalk, the camera could reach sub-electron read noise and combine CCD integration simplicity with CMOS readout speed.","For critically sampled imaging, the sensor's intrinsic point spread function of roughly 1.5 pixels FWHM will broaden stellar PSFs and should be included in PSF fitting or deconvolution."],"supporting_citations":[{"why":"Describes the COSMOS-66 camera architecture, including the 5T backside-illuminated design and dual conversion gain, which the paper characterizes.","marker":"[7]"},{"why":"Supplies the manufacturer's specification sheet used as the comparison baseline for the measured performance.","marker":"[8]"},{"why":"Provides characterization data for other sCMOS sensors (HWK4123, GSENSE 400BSI, IMX455) that populate the comparison table.","marker":"[13]"},{"why":"Documents the 1 m telescope used for the on-sky demonstration and its instrumentation.","marker":"[24]"},{"why":"Measures the 3.69 eV electron-hole pair creation energy in silicon, the value the X-ray conversion gain analysis assumes.","marker":"[31]"},{"why":"Establishes how interpixel capacitance biases conversion gain and quantum efficiency, the basis for the paper's IPC correction.","marker":"[32]"},{"why":"Supplies the stellar spectra used to compute expected fluxes in the photometric validation.","marker":"[39]"}],"fun_headline_variants":["66MP CMOS achieves CCD-class noise after calibration","Per-pixel fix brings 66MP CMOS to 2.9 e− read noise","COSMOS-66: large-format CMOS rivals CCD performance","Pixel-by-pixel correction enables 66MP CMOS for astronomy","66-megapixel CMOS sensor matches CCD noise after fix"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The electron-counting numbers all rest on the assumption that each absorbed X-ray creates one electron-hole pair per 3.69 eV and that the entire charge is collected in the pixel where it landed.","fun_headline_variants_meta":{"raw":{"variants":["66MP CMOS achieves CCD-class noise after calibration","Per-pixel fix brings 66MP CMOS to 2.9 e− read noise","COSMOS-66: large-format CMOS rivals CCD performance","Pixel-by-pixel correction enables 66MP CMOS for astronomy","66-megapixel CMOS sensor matches CCD noise after fix"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000432,"raw_usage":{"total_tokens":2322,"prompt_tokens":1181,"completion_tokens":1141,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":797,"completion_tokens_details":{"reasoning_tokens":1052}},"tokens_in":797,"tokens_out":1141,"duration_ms":11530,"temperature":1.0,"reasoning_tokens":1052,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T20:11:06.560254+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the X-ray-derived conversion gain with a photodiode-calibrated photon flux at a known wavelength and recompute read noise and dark current; if the two gains disagree beyond the quoted errors, all electron-unit values shift in proportion.","supporting_citations":[{"cited_title":"Cheriyan, J","cited_arxiv_id":null,"evidence_quote":"Describes the COSMOS-66 camera architecture, including the 5T backside-illuminated design and dual conversion gain, which the paper characterizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the manufacturer's specification sheet used as the comparison baseline for the measured performance."},{"cited_title":"Khandelwal, S","cited_arxiv_id":null,"evidence_quote":"Provides characterization data for other sCMOS sensors (HWK4123, GSENSE 400BSI, IMX455) that populate the comparison table."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the 1 m telescope used for the on-sky demonstration and its instrumentation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Measures the 3.69 eV electron-hole pair creation energy in silicon, the value the X-ray conversion gain analysis assumes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes how interpixel capacitance biases conversion gain and quantum efficiency, the basis for the paper's IPC correction."},{"cited_title":"Vallenari , A","cited_arxiv_id":null,"evidence_quote":"Supplies the stellar spectra used to compute expected fluxes in the photometric validation."}],"review_version":1}