{"id":"b3818391-e81d-4b91-9b22-1c86dd242635","arxiv_id":"2505.06228","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The cecilia machine learning pipeline recovers abundances for five polluted helium-atmosphere white dwarfs and finds CI-chondrite-like pollutant compositions, including oxygen excesses in two systems.","lead":"Researchers applied a machine learning pipeline called cecilia to five white dwarf stars with helium atmospheres, measuring the chemical elements that fell onto them from destroyed planets. The results suggest the rocky debris is similar to primitive meteorites, with possible oxygen-rich material around two stars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Oxygen-excess claim is fragile: under a Fe2O3 oxidation state SDSS J0859+5732 falls below the 2σ threshold, and using the independent DF12 Fe abundance for SDSS J2311−0041 would reduce its excess; the abstract presents both as robust.","rationale":"I considered two candidate load-bearing concerns. The reader's weakest assumption (He I collisional broadening below 16,000 K) is real but partly mitigated by clipping the strongest He lines, freezing Teff/log g, and validating against DF12 (which uses the same atmosphere grid). The oxygen excess is different: the paper's own numbers show it can fall below significance under Fe2O3 for one object, and the Fe abundance used for J2311 differs by 0.32 dex from DF12, which would reduce the excess. Since the abstract presents this as a central finding without caveats, this is the least secure part of the central claim. It is testable with a straightforward recalculation and does not require new data. The reader mentioned this as a secondary reason for CONDITIONAL but focused the weakest assumption on He broadening; I partially agree. The verdict should remain CONDITIONAL: the paper has merit, but the abstract and possibly the interpretation should be qualified pending the recomputation.","tokens_in":38957,"tokens_out":10157,"duration_ms":89258,"concrete_test":"Recompute the oxygen-excess Monte Carlo for both WDs under four scenarios: (a) cecilia abundances with FeO; (b) cecilia abundances with Fe2O3; (c) DF12 abundances with FeO; (d) DF12 abundances with Fe2O3. Also run a mixed case for J2311 with cecilia's O, Mg, Si, Ca and the DF12 Fe abundance. If the fraction of positive-excess samples falls to ≤95.45% (the Brouwers et al. 2022 2σ threshold) for either object in any scenario, the abstract's 'statistically significant (>2σ)' phrasing must be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's most novel result is that SDSS J0859+5732 and SDSS J2311−0041 show >2σ oxygen excesses. This claim is computed assuming steady-state accretion and FeO, following Rogers et al. (2024b). The paper's own Section 6.1 shows the fragility: with Fe2O3, the positive-excess fractions drop to 94.1% (J0859) and 95.6% (J2311), so J0859 falls below the 95.45% 2σ threshold. Additionally, the cecilia Fe abundance for J2311 is -7.09 dex, while the independent DF12 fit gives -6.77 dex (Table B2), a 0.32 dex difference. A higher Fe abundance binds more O, lowering the excess. Because the abstract states the >2σ result without mentioning the FeO assumption or the DF12 discrepancy, the headline claim is not robust to plausible alternative assumptions. This is load-bearing because the oxygen-rich interpretation depends on it; if the excess is not significant, the main scientific finding is reduced to a tentative hint.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents the first application of the machine-learning spectral pipeline cecilia to five helium-atmosphere polluted white dwarfs (SDSS J0231+2512, J0859+5732, J1109+1318, J1333+6364, and J2311−0041) with no previously well-measured elemental abundances. The authors perform joint Bayesian MCMC fits to SDSS and Keck/ESI spectra, introducing linear-space abundance sampling, per-instrument jitter terms, and a multi-spectrum likelihood. They detect between two and six elements per star, find the pollutants broadly consistent with CI chondrites within 1–2σ, and report >2σ oxygen excesses for SDSS J0859+5732 and SDSS J2311−0041 under a steady-state accretion model with iron in the form of FeO. The abundances are cross-checked against a classical line-by-line fit (DF12) with nominal agreement around 0.2 dex.","tokens_in":39198,"tokens_out":5859,"duration_ms":54785,"significance":"If the results hold, this work demonstrates a practical ML alternative to classical white-dwarf abundance fitting, with the advantage of speed and automatic multi-spectrum fitting. The cross-check against the independent DF12 method, the explicit 0.20 dex systematic floor, the public Keck/ESI and SDSS data, and the reproducibility of the MCMC setup are clear strengths. The claimed oxygen excesses would add two objects to a short list of bodies with evidence for oxygen-rich exoplanetary material. However, the significance of this headline result hinges on the FeO oxidation-state assumption and on the adopted 0.20 dex floor; Section 6.1 itself shows that one of the two excesses drops below the 2σ threshold when Fe2O3 is assumed. The paper's broad 'consistent within 0.2 dex' validation claim is also not fully supported by Table B2. These issues do not invalidate the methodological contribution but require revision before the scientific conclusions can be accepted as stated.","major_comments":[{"comment":"The headline claim of statistically significant (>2σ) oxygen excesses for both SDSS J0859+5732 and SDSS J2311−0041 is not robust to the assumed Fe oxidation state. Section 6.1 itself reports that under the Fe2O3 scenario the positive-excess fractions are 94.1% and 95.6%, so SDSS J0859+5732 falls below the 2σ threshold of 95.45% that the paper adopts from Brouwers et al. (2022). The abstract presents the >2σ result without this FeO assumption or the Fe2O3 caveat. Because the oxygen-rich interpretation is the paper's most distinctive scientific finding, the abstract and Section 6.1 should present the FeO-assumed numbers as such, and ideally state the Fe2O3 numbers in the same breath or replace the '>2σ' language with a more precise statement.","section":"Abstract; Section 6.1"},{"comment":"The validation statement that 'DF12 and cecilia abundances are consistent within 0.2 dex' is contradicted by entries in Table B2. For SDSS J2311−0041, log10(Fe/He) is −6.77 (DF12) versus −7.09 (cecilia), a 0.32 dex difference; for SDSS J1333+6364, log10(O/He) is −5.49 (DF12) versus −5.85 (cecilia), a 0.36 dex difference. These are exactly the elements that matter for the oxygen-excess claim, since a higher Fe abundance binds more oxygen. The validation claim needs to be quantified as e.g. 'mostly within 0.2 dex, with Fe and O discrepancies up to 0.3–0.4 dex,' and the consequences for the oxygen-excess significance should be discussed.","section":"Section 5.1; Table B2"},{"comment":"The error budget for the oxygen-excess significance may be underestimated because Teff and log g are frozen to photometric values and the strongest He I lines are clipped, while the underlying atmosphere models are acknowledged to be inaccurate for neutral-He collisional broadening below 16,000 K. The paper does not propagate the photometric uncertainties in Teff and log g into the derived abundances, and the adopted 0.20 dex noise floor is the only systematic term. Since all five targets lie in the problematic temperature regime, I request a sensitivity test that re-fits at least one target with Teff and log g varied by their photometric ±1σ values, and a statement of how much the retrieved O, Fe, Mg, and Si abundances and the oxygen-excess fraction move. If the shifts are comparable to or larger than 0.20 dex, the reported significance should be reduced accordingly.","section":"Section 3.3; Section 4.3; Section 6.1"}],"minor_comments":[{"comment":"The detectability criterion is stated as σstat,MCMC ≤ 0.10 dex, but all reported total errors are replaced by the 0.20 dex floor; please clarify whether the detection cut is applied to the pre-floor statistical uncertainty and how sensitive the detection list is to using σtot instead.","section":"Section 4.3"},{"comment":"The assumption of a fixed resolving power R=2,000 for SDSS spectra is acknowledged as a limitation in the discussion, but it would be helpful to state in the main text that the actual variation between R≈1,500 and R≈2,500 is not included in the error budget.","section":"Section 3.2.1; Section 6.2"},{"comment":"Several typographical errors remain, including 'oxgyen' and 'enguflment' in Section 6.1/4.4 and 'iwth' in Section 4.3; these should be corrected in proof.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is a real first application of the cecilia ML pipeline to five helium-atmosphere white dwarfs with new Keck/ESI data, and the abundances are checked against a classical line-by-line fit (DF12) to within 0.2 dex. That cross-check is the core contribution—it shows the pipeline works on real DBZ spectra in a regime where the training models are known to be shaky. The authors handle that shakiness about as well as possible: they clip the strong He I lines, freeze Teff and log g to photometric values, add per-instrument jitter, and sample in linear space. They also explicitly say cecilia cannot beat the underlying atmosphere models. The bulk CI-chondrite consistency is not a surprise, but the results are honestly presented.\n\nThe load-bearing novel result is the oxygen excess in SDSS J0859+5732 and SDSS J2311−0041, and that claim is fragile. Section 6.1 shows that under Fe2O3 rather than FeO, the positive-excess fraction for J0859 drops to 94.1%, below the 95.45% 2σ threshold. The independent DF12 Fe abundance for J2311 is 0.32 dex higher than cecilia's, which would bind more oxygen and reduce the excess. The abstract states both as >2σ without mentioning the FeO assumption or the tension with DF12. That is overselling a tentative finding. The sample selection also matters: these five were chosen because they had the most interesting spectral features, so any population-level comment should be taken as illustrative, not statistical.\n\nMy overall take: the abundance measurements and the pipeline validation deserve publication. The oxygen excess claims need to be softened, the abstract rewritten to include the oxidation-state caveat, and the DF12 Fe discrepancy discussed explicitly. A serious referee should not desk-reject this—it has real data, a reproducible method, and a meaningful validation. But the authors should be asked to make the weak points visible in the abstract, not buried in the discussion.","headline":"The ML abundance pipeline is validated on real spectra and the cross-check is credible, but the headline oxygen excesses are more fragile than the abstract admits and should be reined in.","tokens_in":39783,"tokens_out":1446,"would_cite":true,"duration_ms":16779,"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 machine-learning pipeline cecilia extracts two to six elemental abundances from each of five polluted helium-atmosphere white dwarfs, matching classical methods, and finds CI-chondrite-like rocky debris with two significant oxygen…","keywords":["white dwarfs","polluted white dwarfs","exoplanetary composition","machine learning","spectral analysis","helium atmosphere","oxygen excess","CI chondrites"],"falsifier":"Re-analyze the same Keck/ESI spectra for SDSS J0859+5732 and SDSS J2311−0041 using an independent model grid with updated neutral-helium collisional broadening tables and leave $T_{\\mathrm{eff}}$ and $\\log g$ free, then recompute the oxygen excess; if it drops below $2\\sigma$ or the O/Mg ratios shift by more than 0.2 dex, the oxygen-rich interpretation fails. Independently, obtain high-resolution ultraviolet spectra for these two stars to measure O, Si, and Fe without optical blending; the ultraviolet abundances should reproduce the reported excess.","tokens_in":38752,"feed_emoji":"🪐","tokens_out":5907,"duration_ms":53582,"temperature":0.7,"pith_summary":"This paper tests the machine-learning pipeline cecilia on real data: five metal-polluted helium-atmosphere white dwarfs with no previously measured elemental abundances. Fitting SDSS and Keck/ESI optical spectra jointly, cecilia determines the abundances of two to six elements per star with a reported accuracy of about 0.20 dex, on par with conventional white-dwarf atmosphere fitting. The recovered pollutant compositions are broadly CI-chondrite-like, and two stars (SDSS J0859+5732 and SDSS J2311−0041) show oxygen excesses at $>2\\sigma$ significance, suggesting accretion of oxygen-rich exoplanetary material. The paper also upgrades cecilia to sample abundances in linear space and to fit multiple spectra with per-instrument jitter, steps toward population-scale studies.","feed_headline":"Machine learning recovers planet debris chemistry in five white dwarfs","feed_subtitle":"Automated Bayesian fits match classical precision and flag two oxygen-rich polluters.","key_machinery":"The central object is cecilia, a deep neural-network spectral interpolator that turns 13 stellar labels—effective temperature, surface gravity, hydrogen abundance, and ten metal abundances—into a synthetic optical spectrum. Trained on more than 22,000 synthetic spectra from a standard LTE helium-atmosphere code, it predicts a spectrum in about 0.2 seconds, then a Bayesian MCMC with a modified log-likelihood fits the labels. The key upgrades are sampling elemental abundances in linear space (avoiding divergence for non-detections) and adding per-instrument jitter terms so SDSS and Keck/ESI spectra can be fitted jointly. The pipeline's output is then converted into pollutant composition using white-dwarf diffusion timescales and either build-up or steady-state accretion equations.","core_discovery":"The paper's central claim is that cecilia—a neural-network spectral interpolator trained on synthetic helium-atmosphere models—can replace slow 'human-in-the-loop' abundance analysis for polluted white dwarfs without loss of accuracy. Applied to five DZ/DBZ stars, it reports 2, 6, 5, 4, and 6 detected elements respectively, with the two richest systems (SDSS J0859+5732 and SDSS J2311−0041) each showing O, Mg, Si, Ca, and Fe. The pollutant abundances, normalized to Mg and corrected for diffusion in either build-up or steady-state accretion, match CI chondrites within 1–2$\\sigma$. For those two stars the oxygen excess—computed after allocating O to CaO, SiO$_2$, FeO, and MgO—is positive in 95.5% and 96.4% of Monte Carlo samples, crossing the adopted $2\\sigma$ threshold. The authors regard this as evidence that oxygen-rich exoplanetary material can survive post-main-sequence evolution and accrete onto white dwarfs.","pith_inferences":["If the oxygen excesses survive independent analysis, they imply that the parent bodies were hydrated or ice-rich, and that water can be delivered to white dwarfs without being lost during the giant phases; this would strengthen the connection between polluted white dwarfs and water delivery to habitable-zone planets.","The paper's fixed $T_{\\mathrm{eff}}$ / $\\log g$ strategy could be tested by applying cecilia to warm ($T_{\\mathrm{eff}} > 16{,}000$ K) DBZ stars with fully reliable helium broadening; agreement would bracket the systematic error and quantify how much of the reported abundances is model-limited.","The same joint-fit machinery could be extended to DA (hydrogen-atmosphere) white dwarfs, where diffusion timescales are much shorter, to probe recent accretion events with complementary sensitivity.","A direct falsifier: re-derive abundances with an independent non-LTE model for the same ESI spectra and see whether the O/Mg ratios and the Mg I 4481 Å line discrepancy move the oxygen excess below $2\\sigma$."],"forward_implications":["If cecilia's accuracy holds, large spectroscopic surveys (SDSS-V, DESI, WEAVE) can be mined for polluted white dwarfs without manual fitting.","The method recovers compositions that are largely CI-chondrite-like, suggesting that the building blocks of rocky planets in these systems resemble the most primitive Solar System meteorites.","Two systems with $>2\\sigma$ oxygen excess add to the small sample of white dwarfs whose accreted material is water- or oxygen-rich.","The linear-space MCMC and jitter terms show a path toward statistically robust upper limits on undetected elements.","The reported 0.20 dex floor means systematic model uncertainties, not statistical precision, currently limit this class of measurement."],"supporting_citations":[{"why":"Introduces the cecilia pipeline, its neural-network architecture, training set, and baseline predictive accuracy on synthetic data.","marker":"Badenas-Agusti et al. 2024"},{"why":"Provides the LTE helium-atmosphere model code used to generate cecilia's training spectra.","marker":"Dufour et al. 2007"},{"why":"Supplies improvements to helium-rich atmosphere models incorporated into the training grid.","marker":"Blouin et al. 2018a,b"},{"why":"Identified the five targets as He-rich polluted white dwarfs in SDSS and provided the original sample.","marker":"Koester & Kepler 2015"},{"why":"Gives the photometric $T_{\\mathrm{eff}}$, $\\log g$, H, and Ca values used as initial guesses and for freezing those parameters.","marker":"Coutu et al. 2019"},{"why":"Supplies the CI chondrite abundance reference against which the pollutant compositions are compared.","marker":"Alexander 2019a,b"},{"why":"Provides chondritic abundance ratios used to set priors and condensation-temperature ordering.","marker":"Lodders 2003"},{"why":"Defines the Monte Carlo methodology used to compute the oxygen excess and its significance.","marker":"Rogers et al. 2024b"},{"why":"Establishes the $2\\sigma$ significance threshold used to judge the oxygen excess detections.","marker":"Brouwers et al. 2022"}],"fun_headline_variants":["ML pipeline cecilia decodes planet debris on five white dwarfs","Neural network matches classic WD fits, finds oxygen-rich debris","First ML compositional study of five white dwarfs finds oxygen-rich polluters","AI-driven spectral analysis reveals planet debris composition on white dwarfs","Machine learning spots oxygen-rich exoplanet remains on two white dwarfs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results depend on helium-atmosphere models whose collisional broadening of neutral helium lines is known to be inaccurate below about 16,000 K, and all five target stars sit in that regime; the paper clips the strongest helium lines and freezes temperature and gravity, but residual line wings and the same model deficiencies still shape the predicted metal abundances.","fun_headline_variants_meta":{"raw":{"variants":["ML pipeline cecilia decodes planet debris on five white dwarfs","Neural network matches classic WD fits, finds oxygen-rich debris","First ML compositional study of five white dwarfs finds oxygen-rich polluters","AI-driven spectral analysis reveals planet debris composition on white dwarfs","Machine learning spots oxygen-rich exoplanet remains on two white dwarfs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001306,"raw_usage":{"total_tokens":5404,"prompt_tokens":1106,"completion_tokens":4298,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":722,"completion_tokens_details":{"reasoning_tokens":4206}},"tokens_in":722,"tokens_out":4298,"duration_ms":33084,"temperature":1.0,"reasoning_tokens":4206,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:45:27.202844+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-analyze the same Keck/ESI spectra for SDSS J0859+5732 and SDSS J2311−0041 using an independent model grid with updated neutral-helium collisional broadening tables and leave $T_{\\mathrm{eff}}$ and $\\log g$ free, then recompute the oxygen excess; if it drops below $2\\sigma$ or the O/Mg ratios shift by more than 0.2 dex, the oxygen-rich interpretation fails. Independently, obtain high-resolution ultraviolet spectra for these two stars to measure O, Si, and Fe without optical blending; the ultraviolet abundances should reproduce the reported excess.","supporting_citations":[],"review_version":1}