{"id":"73a584b9-61ae-42f2-ab90-c1f04bf5f1c4","arxiv_id":"2412.15324","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Replacing a standard cooling table with an XGBoost model in an NGC300-like isolated galaxy simulation shifts gas temperatures and creates up to 10-20% differences in some C II emission rates.","lead":"This paper runs two versions of an isolated galaxy simulation that differ only in how gas cooling and heating rates are computed, one using a standard interpolation table and one using machine learning. The two runs produce measurably different gas temperatures, phase structures, and carbon emission rates, showing that this modeling choice matters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 5 Myr convergence check in Appendix A does not establish a steady state: thermal timescales for the low-density gas where the main difference is claimed exceed 5 Myr by orders of magnitude, so the phase-diagram differences may be transient.","rationale":"The reader's weakest_assumption already identifies the 5 Myr convergence choice as load-bearing, and I agree. My stress-test sharpens the objection: the Appendix A metric is not a valid steady-state test (the claimed bound on δ is algebraically false for sign-changing residuals), and a simple timescale estimate shows that diffuse gas in the density range of the headline phase-diagram difference cannot have thermally relaxed within 5 Myr. This does not invalidate the controlled one-to-one replacement design, which is a real strength, nor does it prove the differences are artifacts—it means the 5 Myr snapshot is insufficient evidence for the claim that approximation choice alone shifts the thermal state in that regime. Because the issue is addressable by a longer run or a timescale analysis, the appropriate outcome remains CONDITIONAL, matching the reader's verdict rather than moving to accept or reject.","tokens_in":12950,"tokens_out":10033,"duration_ms":66927,"concrete_test":"Compute t_therm = (3/2) k_B T / (n_b |Γ_GH12 − Λ_GH12|) and the corresponding quantity with XGB rates in every (n_b, T) bin of Fig. 1, and overlay contours of t_therm = 5 Myr; then rerun both simulations for 50 Myr and recompute the bottom panel of Fig. 1 and Fig. 2. If the low-density median-temperature ratio or the critical curve moves systematically between 5 and 50 Myr, or if t_therm > 5 Myr in the claimed −3 ≲ log n_b ≲ −1 band, the 5 Myr results are not converged and the central claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central inference—that swapping only the cooling/heating approximation changes the galaxy's thermal state and CII emissivity—depends on the 5 Myr snapshot being a converged response to the new cooling functions. Appendix A's convergence evidence is not sufficient. First, δ_{a,b} in Eq. (A1) is not necessarily bounded by ±1: if Δ_a and Δ_b have opposite signs, |Δ_b − Δ_a| can exceed |Δ_a + Δ_b|. Second, 'no structure' in the δ map is not a steady-state criterion; uncorrelated increments can occur while the underlying Δ(t) is still drifting. Third, and decisively, the radiative thermal relaxation timescale t_therm = (3/2) k_B T / (n_b |Γ−Λ|) at the densities where the main temperature offset is claimed (−3 ≲ log n_b ≲ −1, T ∼ 10^3–10^4 K) is typically 10^7–10^9 yr with the Cloudy-based rates used here—orders of magnitude longer than 5 Myr. Gas in this regime has barely begun to respond to the replacement; the 'systematically hotter' XGB gas and the critical curve at low density may be transient features inherited from the common starting snapshot, not robust consequences of the approximation. Without a physically motivated convergence test, the central claim is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper compares two approximations of the gas cooling and heating function---the GH12 interpolation table and the XGBoost surrogate models of Robinson et al. (2024), both trained on the same Cloudy calculations---by running two otherwise identical hydrodynamics simulations of an isolated NGC300-like galaxy, started from the same t ≈ 600 Myr snapshot of the Semenov et al. (2021) runs, with radiative transfer switched off, constant photoionization rates, and fixed metallicity Z = 0.3 Z_sun. After 5 Myr, the authors report that (i) the gas in the XGB run is systematically hotter at −3 ≲ log n_b ≲ −1, (ii) the phase-diagram mass residual between the runs has a well-defined 'critical curve' of equal gas mass with opposite-signed residuals just above and below it, and (iii) C II excitation-rate ratios r_j deviate from unity by 10–20% at low densities and up to ~40% at high densities. They attribute these differences directly to the approximation choice and connect the critical curve to the different radiative equilibrium temperatures of the two approximations.","tokens_in":13233,"tokens_out":18440,"duration_ms":163720,"significance":"If the reported differences are robust, the paper provides a clean, controlled demonstration that the numerical approximation of the cooling/heating function---not any change in input physics---can shift the predicted thermal state and line-emission rates of simulated gas. The experimental design is a genuine strength: identical initial snapshot, same code, and a single-variable replacement make the causal attribution to the approximation clean. The paper is also commendably explicit about the in-sample nature of the Table 2 accuracy comparison and about the idealized setup (no radiative transfer, fixed metallicity, rate ratios rather than luminosities), and it ships analysis code. The main quantitative claims, however, rest on one pair of runs and one 5 Myr snapshot, and the convergence and significance support provided is not yet strong enough to establish that the reported differences are steady-state consequences of the approximation rather than transient or noise-dominated features.","major_comments":[{"comment":"The convergence argument in Appendix A does not establish that the 5 Myr snapshot reflects a settled response to the replaced cooling and heating functions, so the central attribution claim in Sections 3.1–3.2 is not yet supported. Three specific problems: (i) the statement that the denominator of Eq. (A1) 'ensure[s] that −1 ≤ δ ≤ 1' is false whenever Δ_a and Δ_b have opposite signs; for example, Δ_a = 0.2 with Δ_b = −0.1 gives δ = −3, and Δ_b = −Δ_a makes the denominator vanish, so bins in Fig. 5 that flip sign between snapshots would saturate or overflow the color scale. (ii) An unstructured map of δ measures only the absence of spatial correlation in the incremental change between consecutive snapshots; a slowly drifting residual with uncorrelated per-bin increments would look the same, so 'no structure' is not a stationarity test. (iii) The paper offers no physical timescale argument: for the gas at −3 ≲ log n_b ≲ −1 and T ~ 10^3–10^4 K where the headline temperature offset is claimed, the radiative relaxation time t_therm = (3/2)k_B T / (n_b |Γ − Λ|), with net cooling rates of order 10^-26–10^-24 erg cm^3 s^-1 implied by the Cloudy-based models in this regime, is roughly 10^6–10^9 yr, comparable to or far longer than the 5 Myr elapsed. The 'systematically hotter' XGB gas and the critical curve at low density may therefore be transient features carried over from the common starting snapshot. The authors should either extend the runs to several ISM dynamical times and show the residual map stops evolving, tabulate per-bin t_therm estimates demonstrating 5 Myr ≫ t_therm in the affected density range, or explicitly rescope the conclusions to short-term differences.","section":"Section 2.1 / Appendix A (Eq. A1, Fig. 5)"},{"comment":"The phase-diagram structures that carry the paper's claims, namely the low-density median-temperature offset, the critical curve, and the residual bands around it, are presented without any estimate of their statistical significance. The analysis uses a single pair of runs and one snapshot per run; regions of Fig. 2 are described as 'noise with no clear structure,' but no information is given about how many resolution elements or how much gas mass sit in the bins that define the critical curve, so the zero-crossing contour could be partly a sampling artifact. Similarly, the claim that the XGB run is 'systematically hotter' for −3 ≲ log n_b ≲ −1 is based on the median curves, yet the paper does not report whether the median offset exceeds the inter-run percentile overlap in that density range, a check it does apply to the higher-density range (−1 ≲ log n_b ≲ 1). I recommend reporting the gas mass or cell count per phase-diagram bin and/or bootstrapping the residual maps over cells, so the reader can assess whether the critical curve and the temperature offset are significant rather than consistent with Poisson fluctuations.","section":"Section 3.1 / Figs. 1–2"},{"comment":"The interpretation of the C II comparison is slightly stronger than the calculation supports. The ratio r_j in Eq. (3) is a ratio of emission rates per C II ion, computed with the same temperature-dependent rate coefficients in both runs and with the C II abundance implicitly held fixed. Because the two runs differ in thermal state, and because the cooling/heating models encode different ionization states from the same Cloudy training data, the C II abundance distribution is not guaranteed to be the same in both runs. The measured 10–20% variations in r_j are well defined as stated, but the closing sentence of Section 3.2 ('we would expect the actual C II luminosity to be different') should be qualified: an actual luminosity comparison requires the ionic abundance field, which is not tracked in these runs.","section":"Section 3.2 / Eq. (3)"}],"minor_comments":[{"comment":"The integrals in Eq. (3) are written over dT alone, but r_j is plotted as a function of n_b in Fig. 4; the density conditioning should be written explicitly (for example, the phase-diagram distribution evaluated at fixed n_b and integrated over T).","section":"Eq. (3), Section 2.3"},{"comment":"There are several wording and typographical issues, including 'uses machine learning for the interpolation instead on an analytic function' in the abstract and 'The machine learning models in of Robinson et al. (2024)' in Section 2.2; these should be corrected.","section":"Abstract and Section 2.2"},{"comment":"The in-sample caveat for the mean-squared-error comparison appears in the text but not in the Table 2 caption; moving it to the caption would prevent readers from treating Table 2 as an out-of-sample accuracy claim.","section":"Table 2, Section 2.2"},{"comment":"Bins in Fig. 5 where Δ_a and Δ_b have opposite signs can produce |δ| > 1 or divergent values (see major comment 1); the paper should state explicitly how such bins are handled in the color scale of the δ maps.","section":"Appendix A, Fig. 5"},{"comment":"The equilibrium-temperature explanation for the critical curve assumes that gas at a given density is radiatively relaxed; in a turbulent disk most gas is not at radiative equilibrium, so this mechanism should be labeled as a heuristic interpretation rather than a derived result.","section":"Section 3.1, Fig. 3"},{"comment":"The repository URL in the Acknowledgments ('ngc300 analysis') contains a space and is not a valid URL; the correct link should be provided.","section":"Acknowledgments"}],"recommendation":"major_revision","confidential_remarks":"The manuscript leans largely on the authors' own prior work (GH12, Robinson et al. 2024, Semenov et al. 2021), which is not a problem per se, but it means the accuracy comparison and the simulation infrastructure are not independent checks of each other. The stress-test concern about the 5 Myr convergence genuinely lands: the Appendix A criterion is mathematically flawed in its claimed bound and physically inadequate as a steady-state test, and the radiative thermal timescale in the low-density regime where the headline offset is claimed is plausibly orders of magnitude longer than the run duration. The fixed-Z, no-radiative-transfer setup makes this a narrow idealized test, which is acceptable for a methods-focused paper if the claims are scoped accordingly; I would support publication after the authors either strengthen the convergence evidence or soften the steady-state interpretation, and after adding noise estimates for the phase-diagram structures."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Robinson et al. compare two cooling/heating approximations in an isolated galaxy simulation: the standard GH12 interpolation table and their own XGBoost ML models. The setup is genuinely controlled—same starting snapshot, same code, only the cooling/heating function replaced—so the phase-diagram differences are plausibly attributable to the approximation, not to other physics. The critical-curve explanation in Fig. 3 is a nice touch, and the C II rate ratios (10-20% for some channels) are a useful quantitative output. The paper is also honest about the in-sample nature of the MSE comparison and about the absence of radiative transfer.\n\nThe soft spot is the convergence argument. The paper claims 5 Myr is enough for the new steady-state phase distribution to settle, but Appendix A doesn't demonstrate that. First, the δ metric in Eq. A1 is not bounded by ±1 as claimed: if Δ_a and Δ_b have opposite signs, |Δ_b - Δ_a| can exceed |Δ_a + Δ_b|. Second, \"no structure\" in the δ map isn't a steady-state criterion; a slow drift can look like noise. Third, and decisively, the radiative thermal relaxation timescale at the densities where the main temperature offset is claimed (−3 ≲ log n_b ≲ −1) is orders of magnitude longer than 5 Myr with these rates. So the low-density gas has barely begun to respond to the new cooling functions; the \"systematically hotter\" XGB gas and the critical curve at low density may be transient features inherited from the common starting snapshot.\n\nThis is not a fatal flaw in the sense of invalidating the entire enterprise—the question is well posed and the controlled setup is the right way to answer it—but the headline result is not yet established. A longer run, or at least a proper convergence test that monitors the thermal timescale, is needed before we can trust the phase-diagram offset. The C II ratios are less affected, since they weight mostly denser gas, but they lack uncertainty estimates.\n\nWho should read this? Anyone choosing between interpolation tables and ML emulators for galaxy simulations, and anyone interpreting C II line-intensity mapping forecasts. It deserves a serious referee: the methodology is sound, the writing is clear, and the convergence issue is addressable in revision. I'd send it to peer review, with a strong recommendation that the authors either extend the run or reframe the claims.","headline":"Controlled comparison of cooling-table vs ML emulator in a galaxy simulation, but the 5 Myr convergence check is too weak to support the headline phase-diagram offset.","tokens_in":13786,"tokens_out":2812,"would_cite":false,"duration_ms":25695,"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 choice of how a simulation approximates gas cooling and heating—interpolation table versus machine learning—changes the simulated galaxy's thermal state and its C II emission rates by 10-20%.","keywords":["gas cooling and heating functions","machine learning emulator","interpolation table","temperature-density phase diagram","C II emission","isolated galaxy simulation","NGC300 analog","XGBoost"],"falsifier":"Run the same two cooling-function models with radiative transfer and spatially varying metallicity for longer than 5 Myr; if the systematic temperature offset at low densities and the critical curve disappear or move dramatically, the claim that approximation choice alone reshapes the thermal state would not generalize.","tokens_in":12742,"feed_emoji":"🌌","tokens_out":6201,"duration_ms":51414,"temperature":0.7,"pith_summary":"This paper asks whether the practical choice of how a simulation computes gas cooling and heating rates—rather than any change in astrophysical physics—can alter the galaxy that comes out of the simulation. It compares two approximations in an idealized isolated-galaxy simulation of an NGC300 analog: a standard polynomial interpolation table of Cloudy photoionization calculations, and a machine-learning (XGBoost) model trained on the same calculations, swapped one-for-one in the hydrodynamic code. The runs differ: the machine-learning run is systematically hotter in low-density gas, the two phase diagrams cross along a critical curve, and integrated C II emission rates shift by 10-20% for several collision channels. If these differences persist in more realistic settings, then approximation error alone is a source of uncertainty in simulated galaxy thermal states and in observable line-emission predictions.","feed_headline":"Choosing the cooling table changes a simulated galaxy's thermal state","feed_subtitle":"Swapping the cooling-rate interpolation shifts gas temperatures and changes C II emission by 10-20%.","key_machinery":"The load-bearing object is the one-to-one replacement of the Gnedin and Hollon (2012) polynomial interpolation table with XGBoost gradient-boosted tree models from Robinson et al. (2024), both fit to the same grid of Cloudy photoionization-equilibrium cooling and heating calculations at fixed metallicity $Z = 0.3\\,Z_\\odot$. The comparison machinery is the temperature-density phase diagram residual $\\Delta = (m_{\\rm GH12} - m_{\\rm XGB})/(m_{\\rm GH12} + m_{\\rm XGB})$, whose zero set defines the critical curve, and the equilibrium-temperature crossings $\\Gamma(T_{\\rm equil}) = \\Lambda(T_{\\rm equil})$ that locate the curve relative to the two models' predicted cooling and heating functions.","core_discovery":"The paper establishes that the choice of cooling and heating function approximation, by itself, changes the thermal state of gas in a simulated galaxy. In the run using the machine-learning approximation, low-density gas with $-3 \\lesssim \\log(n_b/\\mathrm{cm}^{-3}) \\lesssim -1$ is systematically hotter than in the run using the interpolation table. The phase diagrams cross along a critical curve in temperature-density space where both runs hold equal gas mass, with the largest mass differences just above and below that curve. Integrated C II emission rates differ by 10-20% for some excitation channels. The authors state that the net cooling function determines gas temperature, and since these simulations tie star formation efficiency to velocity dispersion, the thermal differences could propagate to star formation and feedback.","pith_inferences":["If the machine-learning models are the more accurate of the two at fixed metallicity, then the low-density temperature offset implies that the interpolation table's errors—including occasional negative cooling predictions—are moving gas to cooler phases than the underlying Cloudy rates would produce.","The critical curve is a portable diagnostic: any new cooling/heating approximation could be located relative to an existing run by computing its equilibrium-temperature curve and checking where its phase diagram crosses old ones.","Because C II is a target for line-intensity mapping at redshifts 3-9, a 10-20% systematic in emission efficiency from the cooling approximation could be a relevant systematic for survey forecasts, although the paper does not quantify this.","The 20x slowdown of direct ML replacement suggests practical hybrid routes—precomputing tables from the ML models or distilling them into faster emulators—could capture the accuracy gain without the cost."],"forward_implications":["In the machine-learning run, low-density gas ($-3 \\lesssim \\log(n_b/\\mathrm{cm}^{-3}) \\lesssim -1$) is systematically hotter than in the interpolation-table run.","The two runs' phase diagrams cross along a critical curve where both simulations contain equal gas mass, and the largest mass differences sit just above and below that curve.","At a given density, the critical curve lies at a temperature between the equilibrium temperatures of the two cooling/heating models, so the offset is traceable to where each model balances cooling against heating.","Integrated C II emission rates differ by 10-20% for several excitation channels, meaning observable line predictions inherit the approximation choice.","Because the simulations tie star formation efficiency to the gas velocity dispersion, the thermal differences can propagate into star formation and feedback in longer or more realistic runs."],"supporting_citations":[{"why":"Defines the interpolation-table approximation and its input parameters that the paper replaces as the baseline cooling/heating model.","marker":"Gnedin and Hollon (2012)"},{"why":"Supplies the machine-learning cooling and heating function models whose predictions drive the alternative simulation run.","marker":"Robinson et al. (2024)"},{"why":"Provides the NGC300 analog isolated-galaxy simulation setup and the settled snapshot used as initial conditions.","marker":"Semenov et al. (2021)"},{"why":"Produces the exact Cloudy photoionization-equilibrium cooling and heating calculations that both approximations are trained on.","marker":"Ferland et al. (1998)"},{"why":"Implements the XGBoost gradient-boosted tree algorithm used for the machine-learning cooling and heating functions.","marker":"Chen and Guestrin (2016)"},{"why":"Provides the C II collision and CMB excitation rate expressions used to compute the emission ratios.","marker":"Draine (2011)"}],"fun_headline_variants":["Cooling model choice alters simulated galaxy's gas temperature","Machine-learning cooling heats low-density gas in galaxy sim","Interpolation method shifts phase diagram and CII emission","How cooling function approximation changes galaxy gas state"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that five million years is enough for the gas to reach its new steady state under the swapped cooling functions, and that turning off radiative transfer while fixing metallicity and photoionization rates isolates the approximation's effect.","fun_headline_variants_meta":{"raw":{"variants":["Cooling model choice alters simulated galaxy's gas temperature","Machine-learning cooling heats low-density gas in galaxy sim","Interpolation method shifts phase diagram and CII emission","How cooling function approximation changes galaxy gas state"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1417,"prompt_tokens":913,"completion_tokens":504,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":443}},"tokens_in":529,"tokens_out":504,"duration_ms":4407,"temperature":1.0,"reasoning_tokens":443,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:31:41.946877+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same two cooling-function models with radiative transfer and spatially varying metallicity for longer than 5 Myr; if the systematic temperature offset at low densities and the critical curve disappear or move dramatically, the claim that approximation choice alone reshapes the thermal state would not generalize.","supporting_citations":[],"review_version":1}