{"id":"16684e6a-35a3-4fd6-8bb8-6831f4568b19","arxiv_id":"2508.12032","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A physics-informed neural network reproduces the dark energy background evolution for five EoS parameterizations; MCMC with Pantheon+ data yields constraints consistent with LambdaCDM, with little speed gain over direct analytic evaluation.","lead":"The paper trains physics-informed neural networks to compute the dark energy density in five equation-of-state models and uses them in an MCMC analysis of Pantheon+ supernova data. All five models stay consistent with a cosmological constant, but the claimed computational advantage is modest for these analytically solvable models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MCMC priors exceed the validated surrogate domain for both w0 and Ωm,0, so part of the quoted posterior volume relies on unvalidated extrapolation; closed-form x_de and analytic matter terms permit a decisive recomputation.","rationale":"The reader's weakest assumption was that the PINN surrogate is accurate over the entire MCMC prior volume, specifically flagging the w0 prior below the training range. My reading finds the same structural problem but identifies a second, arguably more consequential instance: the Ωm,0 prior extends to [0.05,0.60] while the matter surrogate is only validated on [0.1,0.4], and several reported posteriors (ΛCDM, BA, JBP, Linear, Logarithmic) have support above 0.4. This matters because the central claim — that all five parametrizations remain consistent with a cosmological constant and that the PINN introduces no significant bias — depends on the likelihood being unbiased across the full sampled region. If the surrogate is inaccurate in those extrapolated corners, the posterior intervals in Table 1 could shift, potentially changing the 95% consistency statement. The concern is not a disagreement with external consensus; it is an internal inconsistency between the stated prior ranges and the validation grids shown in Figs. 4 and 5. The paper's own text even says the priors 'lie entirely within the valid domain,' which the figures contradict. The good news is that the authors provide closed-form xde for all five EoS and the matter term is analytic, so the test is straightforward and decisive. I therefore keep the reader's CONDITIONAL verdict: the paper should be accepted only after the authors either recompute with exact expressions to show the posteriors are unchanged, or retrain/validate the surrogates over the actual prior box and report the resulting constraints.","tokens_in":16798,"tokens_out":8159,"duration_ms":81665,"concrete_test":"Recompute the five MCMC chains using the exact analytic expressions: xde from Eqs. (6)-(10) and xm=Ωm,0(1+z)^3, with exactly the same priors, covariance matrix, and SNe residuals as in Section 4. Compare the resulting marginal 68% and 95% intervals to Table 1. Also record the fraction of accepted samples with w0<−1.6 or Ωm,0 outside [0.1,0.4]; if this fraction is non-negligible, the surrogate extrapolation is exercised by the posterior. If all intervals shift by less than roughly 0.1σ and the extrapolating samples are a negligible fraction with no impact, the concern is resolved and the central claim stands; otherwise the reported posterior constraints are not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 sets the MCMC priors to 0.55≤h0≤0.85, 0.05≤Ωm,0≤0.60, −2≤w0≤0, −3≤wa≤3, and states these 'lie entirely within the valid domain of the trained PINNs.' This is contradicted by the paper's own validation grids: Fig. 5 reports the matter surrogate trained for Ωm,0∈[0.1,0.4], and Fig. 4 reports the dark-energy surrogate validated for w0∈[−1.6,0], wa∈[−8,3], z∈[0,2.5]. The prior box therefore includes w0∈(−2,−1.6) and Ωm,0∈(0.05,0.1)∪(0.4,0.6), all outside the characterised domain. This is not merely cosmetic: the reported posteriors have support in this excluded region. For example, Table 1 gives Ωm,0=0.408±0.021 for ΛCDM (mean above 0.4), BA Ωm,0=0.339+0.089/−0.088, and JBP Ωm,0=0.288+0.15/−0.079, so 1σ tails extend beyond 0.4. If the surrogates extrapolate poorly there, the likelihood is biased in exactly the region where the MCMC accumulates weight, and the quoted 95% consistency statements could shift. Since the paper supplies closed-form xde for all five EoS and the matter term is analytic, this concern is directly testable; it is an internal inconsistency, not a matter of external consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces a cosmology-informed neural network (CINN) framework for background cosmological inference. A physics-informed neural network is trained as a bundle solution for the dark-energy density factor x_de(z; w0, wa) for five equation-of-state parameterizations (CPL, BA, JBP, Linear-z, Logarithmic-z), and an auxiliary network reproduces the matter term. These surrogates are embedded in a Gaussian likelihood for the Pantheon+SH0ES distance moduli, and MCMC is run over (h0, Omega_m,0, w0, wa, M0). The paper's central claims are that the surrogate introduces negligible bias (distance-modulus error below 0.1 mag out to z=2.5) and that all five dynamical dark energy models remain consistent with a cosmological constant at the 95% credible level, with the tightest constraints from CPL. The analytic forms for x_de are derived in Eqs. (6)-(10), validation diagnostics are presented for the matter net, the Hubble rate, and the distance modulus, and a break-even analysis of computational cost is included.","tokens_in":17172,"tokens_out":7555,"duration_ms":83280,"significance":"If the surrogate accuracy is valid over the entire MCMC prior volume, the paper demonstrates a useful and reusable technique: after a one-time training cost, likelihood evaluations become differentiable, GPU-batched, and independent of repeated ODE integration. The honest break-even analysis (about four repeated runs for CPL) and the explicit acknowledgement that the surrogate is not faster for a single simple run are commendable. The derivation of closed-form dark-energy factors for all five parameterizations and the explicit surrogate-error validation are strengths, as is the broad agreement of the CPL constraints with the Pantheon+ team's analysis. The main weakness is that the validation domain documented in Figs. 4 and 5 is narrower than the MCMC prior box stated in Section 4, so the paper's own evidence does not currently establish the central accuracy premise over the full sampled volume.","major_comments":[{"comment":"The uniform priors in Section 4 are said to 'lie entirely within the valid domain of the trained PINNs,' but this is contradicted by the validation grids reported in Section 5.2. Figure 4 validates the dark-energy surrogate on w0 in [-1.6, 0], while the MCMC prior is w0 in [-2, 0]; Figure 5 trains the matter surrogate on Omega_m,0 in [0.1, 0.4], while the MCMC prior is Omega_m,0 in [0.05, 0.6]. The posterior constraints in Table 1 have support outside these ranges, for example Omega_m,0 = 0.408 +/- 0.021 for LambdaCDM and Omega_m,0 = 0.339 +0.089/-0.088 for BA, so the likelihood is being evaluated in unvalidated extrapolation regions. Even if most posterior mass lies inside the validated box, the claim of domain nesting is false and the extrapolation error is uncharacterized. Because the matter term is analytic, the authors should use the closed form Omega_m,0(1+z)^3 in the likelihood, and should retrain or revalidate the dark-energy PINN over the full w0 prior, or alternatively restrict the priors to the validated domain.","section":"Sec. 4; Sec. 5.2 (Figs. 4-5)"},{"comment":"The accuracy claims for the matter surrogate are internally inconsistent. Section 3.1 states that the matter net 'reaches machine-precision (< 10^-6 MSE)' and is 'accurate to less than 10^-5', while Section 5.2 and Figure 5 report relative errors 'at most a few x 10^-2' over the validation grid. A few percent relative error in x_m(z) is not negligible when propagated into E^2(z), and it is incompatible with the sub-per-mille accuracy claimed elsewhere. The authors should report a single consistent error metric (for example, the maximum relative error over the domain) and propagate that error into the E(z) and distance-modulus budgets.","section":"Sec. 3.1; Sec. 5.2 (Fig. 5)"},{"comment":"The paper overstates the surrogate accuracy relative to its own Figure 4. The text accompanying Fig. 4 says the surrogate reproduces E(z) with 'percent level accuracy' and that the largest deviations reach 'a few percent' or up to '0-6%' at the edges, yet the Conclusions state that 'the surrogate model achieves sub-per-mille precision across the (w0, wa, z) space for all parametrizations.' These statements cannot both be true unless the sub-per-mille claim is restricted to a subregion of the domain. The accuracy claims should be restricted to the region actually validated, or the training/validation density should be improved so that the 0-6% edge errors do not occur within the prior volume.","section":"Sec. 5.2 (Fig. 4); Conclusions"},{"comment":"The universal claim that all parameterizations are consistent with a cosmological constant at the 95% credible level should be demonstrated quantitatively in the joint (w0, wa) plane. For example, the JBP marginal in Table 1, w0 = -0.68 +0.14/-0.10, appears to exclude w0 = -1 at more than 2 sigma if interpreted as roughly Gaussian, so the abstract's blanket statement needs support from the 2D contours explicitly containing (-1, 0) at 95% for each model. If the statement refers only to the joint 2D posterior, that should be stated, since the 1D JBP marginal is in apparent tension with it.","section":"Abstract; Sec. 5.1 (Table 1)"}],"minor_comments":[{"comment":"The phrase 'reusable with different datasets' is missing a noun; it should read 'reusable as a tool' or 'reusable module'.","section":"Abstract"},{"comment":"The sentence referring to 'five different omega_a omega_b CDM assumptions' appears garbled; the intended notation is likely w0 wa CDM or similar.","section":"Sec. 1"},{"comment":"The cross-reference 'solves 16' should refer to Eq. (16) explicitly; as written it is unclear.","section":"Sec. 3.1"},{"comment":"The table header gives H0 in km/s/Mpc, but the values listed are 0.7219 etc.; the table should consistently report either h0 (dimensionless) or H0 in km/s/Mpc with values near 72.2.","section":"Table 1"},{"comment":"The caption says the error is obtained by combining 'the ANN-based matter term with analytic dark-energy factors,' while the text describes the dark-energy term as a PINN surrogate; please clarify which quantity is actually used in Fig. 4.","section":"Sec. 5.2 (Fig. 4 caption)"},{"comment":"There are numerous typographical errors, including 'tolrerance', 'cost effectivemenss', 'anisotorpic', and 'cephied'; a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is reasonable and the analytic dark-energy factors are useful, but the validation-domain/prior mismatch and the inconsistent accuracy claims are load-bearing and need to be fixed before publication. I do not see grounds for rejection: the issues are addressable by retraining or revalidating the surrogates over the full prior box, reporting a consistent error metric, and tempering the accuracy and consistency claims. The authors should also be asked to clarify the relationship to the earlier CINN work of Chantada et al. (2023), since the incremental contribution is currently not crisply delineated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a competent, honest extension of the Chantada et al. bundle-PINN pipeline to five dark-energy EoS forms, with better validation than most ML-inference papers. It has one real internal inconsistency—the MCMC priors extend outside the trained surrogate's validated domain—and it oversells the computational savings for models that have closed-form x_de.\n\nWhat's actually new: applying the bundle-PINN approach to BA, JBP, Linear, and Logarithmic EoS in addition to CPL, plus a dedicated matter surrogate and a clear error-propagation chain (matter net → E(z) → distance modulus). The accuracy checks are real: |ΔE|/E < 1e-4 for representative parameter pairs, and |Δμ| < 0.1 mag to z=2.5. The posterior constraints match the Pantheon+ team's CPL results and the paper correctly concludes no significant preference for dynamical dark energy. The paper is also honest about the lack of speedup for a single run and gives a break-even estimate (~4 runs).\n\nThe main problem is the prior/domain mismatch. Section 4 claims the uniform priors lie entirely within the valid domain of the trained PINNs, but Fig. 4 shows w0 training on [-1.6, 0] while the prior goes to -2, and Fig. 5 shows the matter surrogate trained for Ωm,0 ∈ [0.1, 0.4] while the prior is [0.05, 0.6]. The Table 1 posteriors have support in the excluded region—ΛCDM gives Ωm,0 = 0.408 ± 0.021, above 0.4. So part of the quoted posterior volume is built on unvalidated extrapolation. The stress-test note is right that this is directly testable, because the x_de factors are closed-form (Eqs. 6-10) and the matter term is analytic. In fact, once you notice that, the surrogate looks unnecessary for these specific models: the exact likelihood can be evaluated with the analytic expressions at essentially no cost. The paper's computational comparison is against numerical ODE integration, not against the analytic forms, which makes the cost argument somewhat strawman. The honest framing is that the surrogate is a proof-of-concept for models where no closed form exists (e.g., modified gravity), not a practical speedup here.\n\nMinor issues: no code or trained artifacts are provided, and there are numerous typos, but those don't affect the conclusions. If the authors fix the domain inconsistency (either narrow the priors or extend the training grid) and reframe the computational claim, this becomes a solid contribution. I'd send it to peer review.","headline":"Competent extension of bundle-PINN to five EoS forms with real validation, but prior/domain mismatch and overstated speedup need fixing.","tokens_in":17706,"tokens_out":3870,"would_cite":false,"duration_ms":36374,"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":"A physics-informed neural network can stand in for the Friedmann equation inside a supernova likelihood, and when it does, the Pantheon+ data still point to a cosmological constant.","keywords":["physics-informed neural networks","dark energy equation of state","Pantheon+ supernovae","Markov Chain Monte Carlo","cosmological constant","bundle solution","distance modulus","Friedmann equation"],"falsifier":"Evaluate the published surrogate at $(w_0,w_a,z)=(-1.8,0,1.5)$ and compare its distance modulus with exact quadrature of the Friedmann equation; if $|\\Delta\\mu|$ exceeds about 0.1 mag, the claimed prior coverage fails and the posterior bounds in that region are not trustworthy.","tokens_in":16600,"feed_emoji":"🔭","tokens_out":10557,"duration_ms":104816,"temperature":0.7,"pith_summary":"The paper tries to establish that a physics-informed neural network can replace the usual numerical integration of the Friedmann equation when fitting dark energy models to Type Ia supernova data, and that this replacement does not bias the inferred parameters. The authors train one surrogate per equation-of-state parameterization—CPL, BA, JBP, linear-in-$z$, and logarithmic-in-$z$—so that a single network returns the dark-energy density factor $x_{\\rm de}(z;w_0,w_a)$ across the parameter box. Embedding the surrogates in an MCMC likelihood with the Pantheon+ supernova sample, they report distance-modulus errors below 0.1 mag out to $z=2.5$ and posterior constraints in which all five models keep $w_0=-1$, $w_a=0$ inside the 95% credible region. A reader should care because the result suggests the latest supernova compilation does not require dynamical dark energy, and because a validated, differentiable surrogate makes repeated or expensive cosmological likelihood analyses cheaper.","feed_headline":"Neural-net supernova fits still find dark energy is constant","feed_subtitle":"A neural-network stand-in for the Friedmann equation confirms the Pantheon+ sample still prefers a cosmological constant.","key_machinery":"The load-bearing object is the bundle PINN solution: a network $u_\\phi(z,w_0,w_a)$ trained to satisfy the logarithmic form of the dark-energy continuity equation $du/dz=3[1+w(z)]/(1+z)$ over a box of redshifts and equation-of-state parameters, with $x_{\\rm de}(z)=\\exp(u(z))$ encoding the boundary condition $x_{\\rm de}(0)=1$. This exponential reparametrization is what makes the boundary condition exact and the learning target smooth. The bundle solution carries the argument because every quantity in the likelihood—$E(z)$, the comoving distance, and the distance modulus—is a cheap, differentiable function of the network output, so no Friedmann-equation integration is needed during sampling.","core_discovery":"The paper's core claim is that a bundle-style physics-informed neural network, trained once on the continuity equation for dark energy, can replace numerical ODE integration in an MCMC likelihood without measurably biasing parameter inference. For each of five two-parameter equation-of-state forms, the network learns the dark-energy density factor $x_{\\rm de}(z;w_0,w_a)$ across the whole prior box, while a companion network reproduces the analytic matter factor $\\Omega_{m,0}(1+z)^3$, so the dimensionless Hubble rate $E(z)=\\sqrt{\\Omega_{m,0}(1+z)^3+(1-\\Omega_{m,0})x_{\\rm de}(z)}$ is fully differentiable and batch-evaluable. Validation against the analytic solutions gives fractional errors in $E(z)$ below $10^{-4}$ and distance-modulus bias below 0.1 mag out to $z=2.5$. With the Pantheon+ data and an ensemble MCMC sampler, all five parameterizations return $w_0=-1$, $w_a=0$ inside their 95% credible regions, the tightest from the CPL form ($w_0=-0.85\\pm0.14$, $w_a=-0.36^{+0.53}_{-0.24}$), and the inferred Hubble constant stays near 72 km/s/Mpc in every model. The authors read this as evidence that current supernova data are consistent with a cosmological constant, and that the surrogate's practical payoff is for repeated analyses of the same model or for models with expensive likelihood evaluations.","pith_inferences":["A reader reusing the published surrogate should treat the prior region $w_0<-1.6$ as unvalidated, because the stated prior allows $w_0$ down to $-2$ while the published training grid starts at $w_0=-1.6$; comparing surrogate distance moduli with exact quadrature in that region would settle whether the quoted bounds are affected.","The same bundle surrogate could be dropped into a Hamiltonian Monte Carlo or other gradient-based sampler, where its exact derivatives would remove the need for finite-difference tuning of the likelihood.","The method's clearest testbed is a model class whose background equations are expensive, such as modified gravity or anisotropic cosmologies; a natural extension is to train one bundle for such a model and measure the actual break-even number of runs.","The five parameterizations differ in how strongly $w(z)$ can evolve at high redshift, so their similar posteriors suggest that SN data constrain mostly the low-redshift average of $w(z)$ rather than its functional shape."],"forward_implications":["The Pantheon+ Type Ia supernova sample alone does not prefer any of the five evolving-dark-energy parameterizations over a cosmological constant at 95% confidence.","The trained surrogates can be reused for other datasets whose redshifts lie within the training range $z\\le2.5$ without retraining, making the same network applicable to future SN samples.","For simple models such as CPL, the surrogate only saves wall-clock time after roughly four independent analyses of the same model; for models with expensive background evaluations, the break-even point can drop below a single run.","Because the background is differentiable and GPU-batchable, the surrogate enables gradient-based samplers and large parameter surveys that are impractical with per-sample numerical integration.","All five dynamical-dark-energy fits return an $H_0$ in the late-universe range of about 71.8 to 72.7 km/s/Mpc, so the choice of $w(z)$ parameterization does not change the Hubble-tension picture."],"supporting_citations":[{"why":"Supplies the physics-informed neural network residual-loss method used to train the surrogate.","marker":"[5]"},{"why":"Introduces the bundle-solution scheme that lets one network cover the whole parameter box.","marker":"[9]"},{"why":"The earlier cosmology-informed neural network work whose background-dynamics programme this paper extends and validates.","marker":"[10]"},{"why":"Provides the Cepheid-calibrated supernova distances and local Hubble constant measurement used for comparison.","marker":"[12]"},{"why":"Defines the cosmological-constraints analysis whose calibrated distance moduli are used in the likelihood.","marker":"[13]"},{"why":"Defines the CPL equation-of-state form that yields the paper's tightest constraints.","marker":"[25, 26]"},{"why":"Defines the Barboza-Alcaniz equation-of-state form used as one of the five models.","marker":"[32]"},{"why":"Defines the Jassal-Bagla-Padmanabhan equation-of-state form used as one of the five models.","marker":"[33, 34]"},{"why":"Supplies the supernova sample and covariance matrix used in the likelihood.","marker":"[43]"},{"why":"Provides the ensemble MCMC sampler used to draw the posteriors.","marker":"[47]"}],"fun_headline_variants":["Neural nets confirm dark energy is still constant","Cosmology-informed AI sees no dark energy drift","Supernova data via neural nets: dark energy constant","PINN emulator keeps dark energy at -1, says Pantheon+","AI-driven analysis: dark energy equation of state still lambda"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The inference rests on the assumption that the trained surrogate is accurate over the entire MCMC prior volume, but the prior allows $w_0$ down to $-2$ while the published training grid starts at $w_0=-1.6$, so samples below that edge are extrapolations whose error is uncharacterized.","fun_headline_variants_meta":{"raw":{"variants":["Neural nets confirm dark energy is still constant","Cosmology-informed AI sees no dark energy drift","Supernova data via neural nets: dark energy constant","PINN emulator keeps dark energy at -1, says Pantheon+","AI-driven analysis: dark energy equation of state still lambda"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00044,"raw_usage":{"total_tokens":2346,"prompt_tokens":1172,"completion_tokens":1174,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":788,"completion_tokens_details":{"reasoning_tokens":1093}},"tokens_in":788,"tokens_out":1174,"duration_ms":13214,"temperature":1.0,"reasoning_tokens":1093,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:26:43.741684+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the published surrogate at $(w_0,w_a,z)=(-1.8,0,1.5)$ and compare its distance modulus with exact quadrature of the Friedmann equation; if $|\\Delta\\mu|$ exceeds about 0.1 mag, the claimed prior coverage fails and the posterior bounds in that region are not trustworthy.","supporting_citations":[{"cited_title":"Chantada, Susana J","cited_arxiv_id":null,"evidence_quote":"The earlier cosmology-informed neural network work whose background-dynamics programme this paper extends and validates."},{"cited_title":"The Pantheon+ Analysis: Cosmological Constraints","cited_arxiv_id":null,"evidence_quote":"Defines the cosmological-constraints analysis whose calibrated distance moduli are used in the likelihood."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Barboza-Alcaniz equation-of-state form used as one of the five models."},{"cited_title":"The pantheon+ analysis: the full data set and light-curve release","cited_arxiv_id":null,"evidence_quote":"Supplies the supernova sample and covariance matrix used in the likelihood."},{"cited_title":"Hogg, Dustin Lang, and Jonathan Goodman","cited_arxiv_id":null,"evidence_quote":"Provides the ensemble MCMC sampler used to draw the posteriors."}],"review_version":2}