{"id":"3e7a20fc-513c-467b-a86b-a65624de5f2d","arxiv_id":"2412.03751","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A two-fluid subgrid model with only two tuned parameters reproduces resolved multiphase gas evolution in turbulent boxes, enabling unresolved cold gas in large-scale astrophysical simulations.","lead":"This paper introduces MOGLI, a subgrid model that tracks hot and cold gas as two separate fluids in simulations, with source terms for drag, turbulent mixing, and cold-gas growth. The authors show that MOGLI reproduces resolved simulations of cloud destruction and growth across a wide parameter range, potentially letting large-scale galaxy simulations model cold gas without resolving parsec-scale clouds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Verification is partly circular: the growth/destruction rates and the two tuned parameters are drawn from the same resolved-simulation class used as the benchmark, so the claimed reproduction is partially a fitting result.","rationale":"The reader’s weakest_assumption was the effective-sphere geometry and its area-volume relation (Eq. 13, Eq. 30). That is a real physical approximation, but it is not the most load-bearing for the stated central claim. Within the turbulent-box benchmark class, unresolved clouds genuinely are spherical by construction, and for resolved clouds the alpha_mass threshold (not the sphere geometry) is what prevents interior destruction; the geometry is effectively part of the calibration machinery. The deeper issue is that the model’s rates and its two free parameters are derived from, and tuned to, the same resolved simulations used for verification. This threatens the inference ‘MOGLI reproduces resolved behavior’ because the agreement is partly an artifact of fitting. The reader’s rationale did mention ‘reuse of the same benchmark class for fitting and verification,’ but their chosen weakest assumption did not centre on it, so I mark partial agreement. I keep the verdict CONDITIONAL because the concern does not invalidate the model—it requires an out-of-sample test to establish predictive power. The paper is transparent about the parameters and limitations, and the design is physically motivated, so I do not advocate rejection. The concrete test is a single, feasible check that would settle whether the apparent reproduction is robust beyond the calibration regime.","tokens_in":30493,"tokens_out":8788,"duration_ms":87818,"concrete_test":"Freeze alpha_mass = 0.15 and xi = 2 as published, and run MOGLI on a turbulent box with a density contrast chi = 300 (instead of the calibrated chi = 100), keeping box size, Mach number, and cooling setup matched to the existing benchmarks. Compare the half-mass time (Fig. 7) and growth-rate ratio (Fig. 11) against new resolved Athena++ simulations at equivalent resolution. If MOGLI’s values deviate by more than the 2-sigma scatter of the benchmark set, the model is overfit to the chi = 100 calibration. A complementary test: re-tune the two parameters using only a subset of the calibration suite (e.g., M = 0.5 solved clouds) and then predict the rest; if the optimal parameters move substantially, the model’s apparent parameter independence is not robust.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that MOGLI reproduces resolved single-fluid behavior in turbulent boxes. The load-bearing condition is that this agreement constitutes genuine predictive validation, not in-sample fitting. That condition is insecure. The destruction timescale (Eq. 12) and growth timescale (Eq. 17) are taken directly from Gronke et al. (2022) and Tan et al. (2021), and the benchmark comparison in Fig. 11 explicitly includes those same resolved simulations (plus Das & Gronke 2024). The two free parameters are also tuned to the Athena++ turbulent boxes: alpha_mass = 0.15 is set because it ‘works well across all tests’ (Sec. 3.4), and xi = 2 is chosen during the non-radiative mixing tests (Sec. 5). Thus the agreement in thalf/tcc and tgrow (Figs. 7, 11) is not an independent check of the rates—it measures how well the fitted rates reproduce the data they were fitted to. The survival criterion (Fig. 12) is presented as emergent, but because the per-cell growth and destruction rates were constructed from the same theory that predicts the criterion, its reproduction is a consistency check rather than a genuinely new prediction. The most independent diagnostic, cold-gas dispersion (Fig. 13), shows only qualitative agreement and a systematically higher dispersion in MOGLI, which the authors attribute to numerical diffusion. This does not refute the model, but it shows the quantitative agreement does not extend to all diagnostics. The concern is not that the model is unusable—it is a well-motivated construction—but that the evidence does not yet separate the model’s physical content from the calibration used to build it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Das, Gronke, and Weinberger present MOGLI, a subgrid model for unresolved cold gas implemented in the multifluid extension of AREPO. The model evolves hot and cold fluids with source terms for three physical processes: drag, turbulent mixing of cold into hot gas, and cold-gas growth by cooling of mixed gas. Local turbulent velocities are estimated either by Kolmogorov scaling from the box scale or locally from velocity gradients. The cold-gas surface area and cross-sectional area are represented through Monte-Carlo fits for ensembles of overlapping spheres. The paper verifies both a reduced non-radiative version and the full radiative version of the model against resolved Athena++ turbulent-box simulations, varying Mach number, spatial resolution, initial cloud resolvedness, and turbulence seeds. The reported diagnostics include cold-gas destruction timescales, growth rates, the Gronke et al. (2022) survival criterion, and cold-gas dispersion. The paper closes with a demonstration of a 64^3-cell simulation containing 100 unresolved clouds, which the authors argue would require roughly 3000^3 cells in a resolved single-fluid calculation.","tokens_in":30893,"tokens_out":12702,"duration_ms":133708,"significance":"The model addresses a real and widely recognized need: cosmological and galaxy-scale simulations cannot resolve the small cold clouds observed in galactic halos, and current simulations are non-converged in cold-gas content. The manuscript's main strength is its extensive verification suite: it tests a reduced model and the full model, compares two independent local-turbulence estimators, uses quantitative diagnostics (t_half, t_grow, survival, dispersion) across Mach number, resolution, resolvedness, and random seeds, and contains an honest discussion of limitations. The implementation in a widely used code and the small number of free parameters are additional assets. The main caveats are that the two free parameters are selected using the same benchmark class used for validation, and that the survival criterion is in large part a consistency check with the growth/destruction theory built into the model. If the calibration/validation distinction is addressed, the model would be a valuable tool for future large-scale multiphase simulations.","major_comments":[{"comment":"The two model parameters are calibrated against the same benchmark suite used for validation: alpha_mass=0.15 is introduced because it 'works well across all tests' (Sec. 3.4), and xi=2 is selected during the non-radiative mixing tests by comparing with benchmark Athena++ runs (Sec. 5). Consequently, the agreement in Figs. 7, 11, and 12 is partly in-sample fitting rather than independent prediction. To support the abstract's claim of predictive verification, the paper should provide a sensitivity scan over alpha_mass and xi and/or perform a holdout exercise in which one portion of the parameter space is used for calibration and the remainder for validation. The discussion should also clearly separate calibrated quantities from genuinely out-of-sample predictions.","section":"Secs. 3.4 and 5; Eqs. (15) and (29)"},{"comment":"The per-cell growth timescale in Eq. (17) is built from the same theoretical framework that produces the analytical global growth rate in Eq. (34), and the Monte-Carlo area factor is explicitly introduced to reproduce the 0.5 fudge factor from Gronke et al. (2022). Therefore Fig. 11 and the survival criterion in Fig. 12 verify the internal consistency of the implemented rates rather than an emergent, independent prediction. The paper should either identify a genuinely independent prediction not used in calibration or reframe the survival criterion as a consistency check rather than an emergent result.","section":"Sec. 4.2.2; Eqs. (17) and (34)"},{"comment":"Cold-gas dispersion is a diagnostic directly relevant to the intended applications, and the MOGLI runs overpredict it substantially: in the unresolved initial-cloud case, the normalized dispersion reaches roughly 10^4 at t = 2.5 t_eddy while the Athena++ benchmarks reach roughly 10^2, and the resolved case also shows a systematic offset. The text attributes this difference to numerical diffusion, but the abstract's statement of 'very good quantitative agreement across the different simulation parameters and diagnostics' and the conclusion bullet that dispersion is 'similar' are too strong. Please either reduce the discrepancy or quantify and qualify the claim, reporting absolute as well as normalized dispersion.","section":"Sec. 4.2.3; Figs. 13 and 18"},{"comment":"The interface-area and cross-sectional-area relations are derived from ensembles of overlapping spheres, and the validation benchmarks are initialized with spherical clouds. If the unresolved cold gas is filamentary or has a different area-volume relation, the mixing and growth rates in Eqs. (15) and (18) would change. Since the intended applications include small, potentially elongated cold structures in the CGM, the paper should either provide a sensitivity test with non-spherical initial morphologies or explicitly state this geometric assumption as a central limitation rather than implying general validity.","section":"Sec. 3.7; Eqs. (30) and (33)"}],"minor_comments":[{"comment":"The text describing the right two columns of Fig. 8 labels the resolved-cloud case as L_box/R_cloud = 32, but Sec. 4.1 and the surrounding discussion use L_box/R_cloud = 8 for resolved clouds; this is inconsistent and should be corrected.","section":"Fig. 8 and Sec. 4.2.1"},{"comment":"For a single non-overlapping sphere the geometric limit is A R/(2V) = 1.5 alpha, whereas Eq. (30) gives h=1 in the dilute limit, which corresponds to A R/(2V) = alpha; a sentence clarifying how the Monte-Carlo fit treats overlapping spheres and how the normalization was chosen would help.","section":"Sec. 3.7, Eq. (30)"},{"comment":"The origin of the alpha^(1/9) factor in Eq. (17) is not derived in the text; a one-sentence justification or a reference for this scaling would improve readability.","section":"Sec. 3.5, Eq. (17)"},{"comment":"The quasi-isothermal EOS for the cold fluid, which resets the internal energy after each timestep, is a significant approximation; its effect on the energy-exchanging source terms in Eqs. (11) and (16) should be discussed or tested.","section":"Sec. 2.3"},{"comment":"The derivation of the velocity-gradient estimator assumes a uniform distribution of neighbours and roughly equal neighbour distances; the manuscript would benefit from a quantitative error estimate beyond the factor-of-two statement in the caption of Fig. 3.","section":"Sec. 3.6.2, Eqs. (24)-(27)"},{"comment":"The possible sensitivity of the area and cross-section fits (Sec. 3.7) to non-spherical subgrid morphology is not listed among the limitations; given the intended CGM applications, this deserves explicit mention in the limitations section.","section":"Sec. 6.4"}],"recommendation":"major_revision","confidential_remarks":"The calibration/validation overlap with the benchmark papers is natural given the authors' prior work, but it should be foregrounded in the revised text. I would support publication after the requested revisions: the sensitivity and holdout tests are feasible within the paper's scope, and the dispersion discrepancy can be either reduced or clearly qualified. The manuscript is a useful contribution to a pressing methodological problem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: MOGLI is real new machinery—a compressible second fluid for cold gas with arbitrary volume filling, drag/mixing/growth couplings, and a new velocity-gradient turbulence estimator. That puts it beyond the pressureless two-fluid work and the Eulerian-Lagrangian particle models. The verification suite is extensive: non-radiative and radiative runs, resolved and unresolved clouds, resolution and Mach number variations, multiple turbulence seeds, and quantitative comparisons of t_half, t_grow, dispersion, and the survival boundary. The 100-cloud showcase makes the practical payoff concrete.\n\nThe main soft spot is exactly what the stress-test flags: the agreement is partly a fitting result. The destruction and growth rates (Eqs. 12 and 17) are imported from the same resolved-simulation class used as benchmarks (Gronke et al. 2022, Tan et al. 2021, Das & Gronke 2024), and the two free parameters (alpha_mass = 0.15, xi = 2) were tuned on the same turbulent-box tests. So Figs. 7 and 11 largely measure how well the fitted rates reproduce the data they were fitted to. The survival criterion (Fig. 12) is presented as emergent, but because it follows from the same theory that gave the per-cell rates, it is more a consistency check than an independent prediction. The most independent diagnostic—cold-gas dispersion (Fig. 13)—shows qualitative agreement with systematically higher dispersion in MOGLI, which the authors attribute to numerical diffusion. That does not refute the model, but it keeps the evidence from being fully out-of-sample.\n\nI do not want to overstate the problem. The authors are transparent about what is tuned and what is assumed, and the model is built from physically motivated pieces rather than free-form fitting. The circularity is real but not damning: the rates are not arbitrary, and the model reproduces behavior across a wide parameter range with a minimal number of parameters. For a first presentation of a subgrid model, this is a solid foundation.\n\nWho is this for? Anyone working on CGM/ICM simulations and the non-convergence problem. It deserves a serious referee and publication after revision, but the in-sample nature of the verification should be stated more carefully, and an independent test (a different setup, or a prediction made before further tuning) would strengthen it. I would send it out.","headline":"A genuine new subgrid framework with honest testing, but the verification is partly in-sample; worth a serious referee.","tokens_in":31377,"tokens_out":1432,"would_cite":true,"duration_ms":15180,"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":"MOGLI is a two-fluid subgrid model that reproduces resolved multiphase gas behavior with only two free parameters.","keywords":["multiphase gas","subgrid model","multifluid hydrodynamics","turbulent mixing","cold gas growth","cloud survival","circumgalactic medium","numerical methods"],"falsifier":"Run a high-resolution resolved single-fluid turbulent-box simulation initialized with a cold phase that is filamentary or produced by shattering rather than a single spherical cloud, measure the cold-gas surface area per cell as a function of $\\alpha$, and compare the MOGLI-predicted destruction and growth timescales (Eqs. 15 and 18) with the resolved evolution. A substantial deviation from the overlapping-sphere area fit (Eq. 30) would show the closure is not universal.","tokens_in":30303,"feed_emoji":"☁️","tokens_out":8101,"duration_ms":74072,"temperature":0.7,"pith_summary":"This paper introduces MOGLI, a subgrid model that lets simulations track unresolved cold and hot gas as two fluids occupying the same grid cell, instead of requiring the cold clouds to be resolved. The model's source terms split phase interactions into hydrodynamic drag, turbulent mixing of cold gas into hot gas, and growth of cold gas by cooling of mixed gas, with local turbulent velocity estimated either by Kolmogorov scaling or from the velocity gradient. The central claim is that with only two free parameters this prescription quantitatively matches resolved single-fluid turbulent-box simulations across resolution, Mach number, turbulence realizations, and resolved or unresolved initial clouds, including destruction timescales, growth rates, and the cloud survival criterion. If this holds, the model gives large-scale simulations of galaxy halos and clusters a computationally cheap route to include the parsec-scale cold gas that observations show is ubiquitous.","feed_headline":"Two-free-parameter subgrid model matches resolved multiphase gas","feed_subtitle":"MOGLI tracks hot and cold gas as two fluids, matching destruction, growth, and survival in turbulent boxes.","key_machinery":"The load-bearing machinery is the cell-level source-term closure. Cold gas in a cell is assigned an effective spherical size $l_{\\rm cold}=(\\alpha V_{\\rm cell}/(4\\pi/3))^{1/3}$, and all exchanges between the cold and hot fluids are written as rates proportional to the cold-gas surface area, using the Monte-Carlo fit $2h(\\alpha)$ for overlapping spheres at volume fraction $\\alpha$, times a flux with timescale $t_{\\rm destroy}=\\chi^{1/2} l_{\\rm cold}/v_{\\rm turb}$ for destruction and $t_{\\rm grow}=\\chi(t_{\\rm destroy}t_{\\rm cool,cold})^{1/2}\\alpha^{1/9}$ for growth. Mixing is switched off in cells where the cold mass fraction exceeds $\\alpha_{\\rm mass}=0.15$, encoding shielding of cloud interiors, and the gradient-based turbulence estimator uses $\\xi=2$ as the normalization in Eq. (29). These two values are the only free parameters; everything else, including the cross-sectional drag area $A_{\\rm cross}(\\alpha)$, follows from geometry and earlier resolved-simulation results. The machinery's job is to convert a cell's unresolved cold-gas volume fraction into the rates at which that gas mixes, drags, and grows.","core_discovery":"The paper claims that unresolved multiphase gas can be represented faithfully by two co-located fluids whose interactions are governed by physically motivated rates rather than by tuned destructive behavior. The mass exchange between phases is written as an interface area times a flux, with the area set by a Monte-Carlo fit for overlapping cold spheres at volume fraction $\\alpha$, and the flux timescales set by the cloud-crushing destruction time $\\chi^{1/2} l_{\\rm cold}/v_{\\rm turb}$ and the cooling-regulated growth time $\\chi (t_{\\rm destroy}t_{\\rm cool,cold})^{1/2}\\alpha^{1/9}$. A cold-mass-fraction threshold suppresses mixing in shielded cold-dominated cells, and a gradient-normalization constant calibrates the local turbulence estimate. The paper verifies the full model against resolved single-fluid simulations, finding that cold gas destruction and growth rates, the emergence of the survival criterion from prior resolved-simulation work, and cold gas dispersion all match within the stochastic scatter of the benchmarks. It also demonstrates a 64$^3$-cell run tracking 100 unresolved clouds that would need roughly 3000$^3$ cells in a single-fluid code.","pith_inferences":["If the overlapping-sphere area fit is universal, MOGLI should also capture statistically filamentary subgrid gas, because only the cell-level area-to-volume relation matters; testing with resolved simulations that start from filamentary or shattering-produced geometries would settle this.","The gradient-based velocity estimator could be reused as a general subgrid turbulence proxy in other multi-fluid codes, since it only needs the local velocity Jacobian; the paper notes slope-limiting distorts high-velocity tails, so accuracy in strongly shocked flows is an open question.","The model's shielding threshold ($\\alpha_{\\rm mass}=0.15$) acts like a per-cell switch between destruction and growth; a similar switch might be needed in three-phase extensions (molecular-cold-hot), but the paper leaves that to future work."],"forward_implications":["Cold gas mass in large-scale simulations becomes resolution independent: a 64$^3$ multifluid run tracks what would require roughly a 3000$^3$ single-fluid run to resolve, so cosmological and halo simulations can include subgrid cold gas without resolving parsec-scale clouds.","The survival criterion for cold clouds (the boundary in $t_{\\rm cool}/t_{\\rm cc}$ and Mach number where clouds grow instead of being destroyed) emerges from the local source terms, so it is not an input to the model.","The gradient-based turbulence estimator makes the model applicable where turbulence varies in space and time, not only in boxes with a globally known turbulent velocity.","Growth rates match the analytic prediction $t_{\\rm grow}\\sim 1.5\\,t_{\\rm grow,theory}$ with scatter comparable to resolved simulations, so the model can quantitatively predict cold gas mass evolution in growth-dominated regimes."],"supporting_citations":[{"why":"Establishes the analytical growth timescale and the cloud survival criterion in resolved simulations that MOGLI reproduces as emergent behavior.","marker":"Gronke et al. 2022"},{"why":"Provides the combustion-theory-based entrainment rate that motivates the area-factor mass exchange expression and the $t_{\\rm grow}$ timescale.","marker":"Tan et al. 2021"},{"why":"Supplies resolved radiative turbulent-mixing simulations used as the benchmark for survival, growth, and scatter comparisons.","marker":"Das & Gronke 2024"},{"why":"Provides the multifluid hydrodynamics framework in AREPO that carries the two co-located fluids and source-term integration.","marker":"Weinberger & Hernquist 2023"},{"why":"Gives the canonical cloud-crushing destruction timescale $t_{\\rm destroy}=\\chi^{1/2} l_{\\rm cold}/v_{\\rm turb}$ used for the mixing term.","marker":"Klein et al. 1994"},{"why":"Supplies the CIE cooling curve used in the resolved benchmarks and to convert $t_{\\rm cool,cold}$ to the mixed-gas cooling time in the survival plot.","marker":"Wiersma et al. 2009"},{"why":"The Athena++ code used for the resolved single-fluid benchmark simulations.","marker":"Stone et al. 2020"},{"why":"Derives the neighbour-gradient estimator in AREPO that the grad local-turbulence method is built on.","marker":"Pakmor et al. 2016"}],"fun_headline_variants":["MOGLI: two-fluid model matches resolved cold gas runs","MOGLI: two fluids reproduce cold gas survival and destruction","MOGLI: two-fluid model, emergent survival criterion","MOGLI: 64^3 cells simulate 100 cold clouds","MOGLI: two-parameter subgrid scheme matches resolved benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model's rates depend on the assumption that unresolved cold gas inside a cell occupies the surface area of an effective sphere (or overlapping spheres) at the cell's volume fraction; if the real subgrid cold gas is filamentary, sheet-like, or otherwise has a different area-to-volume relation, the mixing and growth rates would be wrong.","fun_headline_variants_meta":{"raw":{"variants":["MOGLI: two-fluid model matches resolved cold gas runs","MOGLI: two fluids reproduce cold gas survival and destruction","MOGLI: two-fluid model, emergent survival criterion","MOGLI: 64^3 cells simulate 100 cold clouds","MOGLI: two-parameter subgrid scheme matches resolved benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001651,"raw_usage":{"total_tokens":6637,"prompt_tokens":1107,"completion_tokens":5530,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":723,"completion_tokens_details":{"reasoning_tokens":5452}},"tokens_in":723,"tokens_out":5530,"duration_ms":34718,"temperature":1.0,"reasoning_tokens":5452,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:07:52.296112+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a high-resolution resolved single-fluid turbulent-box simulation initialized with a cold phase that is filamentary or produced by shattering rather than a single spherical cloud, measure the cold-gas surface area per cell as a function of $\\alpha$, and compare the MOGLI-predicted destruction and growth timescales (Eqs. 15 and 18) with the resolved evolution. A substantial deviation from the overlapping-sphere area fit (Eq. 30) would show the closure is not universal.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CIE cooling curve used in the resolved benchmarks and to convert $t_{\\rm cool,cold}$ to the mixed-gas cooling time in the survival plot."}],"review_version":1}