Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Emulating compact binary population synthesis simulations with uncertainty quantification and model comparison using Bayesian normalizing flows

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Sampling flow weights with Hamiltonian Monte Carlo gives calibrated error bars for emulated binary populations

desk verdict Useful proof-of-concept for exact Bayesian UQ in flow-based PopSynth emulators, but the test-set-tuned prior width undercuts the headline calibration claim. read the letter →

arxiv 2506.05657 v2 pith:D577LF36 submitted 2025-06-06 astro-ph.HE cs.LGgr-qc

Anarya Ray This is my paper
classification astro-ph.HEcs.LGgr-qc
keywords PopulationsynthesisMachinelearningGravitationalwavesourcesCompactbinarystarsHighenergyastrophysicsnormalizingflowsuncertaintyquantificationHamiltonianMonteCarlo
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Population synthesis simulations map how merging black-hole binaries form from their stellar ancestors, but individual runs are expensive, so astronomers replace them with normalizing flows that learn the conditional distribution of binary properties given astrophysical initial conditions by transforming a simple base distribution through invertible neural-network layers. The paper argues that a single best-fit flow, or even an average of an ensemble of them, misses part of the problem: when the flow is trained on few simulated binaries, here as few as 20 per grid point, its predictions wander between retrainings, and that variance needs to be quantified. The proposed remedy is to treat the flow as a Bayesian neural network and sample the posterior distribution of its parameters using Hamiltonian Monte Carlo with the exact likelihood that comes naturally with any density estimator, then predict from the averaged, posterior-predictive flow. On simulated binary black hole populations built through common envelope evolution, the resulting credible intervals contain the true population at the advertised rate, a property called calibration, and they outperform an ensemble of 150 re-trained flows on the same sparse data. If correct, this gives gravitational-wave astronomy a recipe for trusting and marginalizing over emulator error when inferring astrophysics from growing catalogs and when predicting features of rare sub-populations.

What carries the argument

The load-bearing object is the conditional normalizing flow $\hat{p}(\theta|\lambda,\omega) = p_u\big(f_{\mathrm{NN}}(\theta,\lambda,\omega)\big)\,\big|\det \partial f_{\mathrm{NN}}/\partial\theta\big|$, whose parameters $\omega$ are the weights of the neural-network transformations. The key identity is that the flow's training objective, the KL divergence of Eq. (3), is up to a constant the exact log-likelihood of the flow parameters, so the Bayesian posterior $p(\omega|\mathbf{d}_{\mathrm{train}}) \propto p(\mathbf{d}_{\mathrm{train}}|\omega)\,p(\omega)$ is directly samplable with NUTS, with no approximate loss and no assumption about the noise distribution. The prior is an affine-transformed uniform distribution centered on the maximum-likelihood weights $\omega_0$ with a maximal scale $\sigma_0$, deliberately restrictive for convergence; calibration curves comparing empirical to nominal coverage on the held-out test population select $\sigma_0 = 0.25$. The posterior predictive $\bar{p}_m(\theta|\lambda,\mathbf{d}_{\mathrm{train}}) = \tfrac{1}{M}\sum_{i} \hat{p}(\theta|\lambda,\omega_i)$ replaces the deterministic emulator, and the same machinery yields the BIC used for architecture comparison.

What would settle it

Re-run the pipeline with the prior width chosen without touching the test set, for example a hierarchical prior on $\sigma_0$ or calibration on a validation split only, and check whether coverage stays calibrated on a freshly simulated test population at several grid points, not just $\alpha = 1.0$, $\chi_b = 0.1$. A sharper check: compute coverage per parameter-space bin rather than averaged over bins, focusing on the high-chirp-mass, unequal-mass-ratio tail that motivates data augmentation; if the 90% credible intervals under-cover in exactly that sparse region, or a full $10^6$-binary simulator rerun falls outside them, the calibration claim fails in the regime the method is built for.

Watch

Extended reading notes

Core claim

The central claim is that a normalizing-flow emulator of population synthesis should be treated as a Bayesian model rather than a fitted curve. Because a density estimator carries its own exact likelihood, $\log p(\mathbf{d}_{\mathrm{train}}|\omega) = \sum_i \log \hat{p}(\theta_i|\lambda_i,\omega)$, combining it with a prior on the flow parameters defines a posterior $p(\omega|\mathbf{d}_{\mathrm{train}})$ that can be sampled with the No-U-Turn sampler; prediction then proceeds by averaging over the posterior draws, producing a credible interval for the emulated density instead of a single curve. On two- and four-dimensional populations of merging black holes formed through common-envelope evolution, with training sets thinned to 20 binaries per grid point, the posterior-predictive intervals enclose the true test population in both density evaluation and sample generation, and empirical coverage tracks nominal coverage once the prior scale $\sigma_0 = 0.25$ is adopted. The paper further shows these intervals are better calibrated than an ensemble of 150 maximum-likelihood retrainings, that the same posterior samples support architecture comparison through the Bayesian information criterion, and that uncertainty-carrying data augmentation predicts features of rare sub-populations such as unequal-mass-ratio, high-chirp-mass binaries with intervals that contain the simulator's own large-scale rerun.

Load-bearing premise

The prior on the flow parameters is centered on the maximum-likelihood fit computed from the same training data, and its width is chosen by tuning against the held-out test set's calibration curve, so if that prior is too narrow or the tuning overfits the test set, the quoted coverage is overconfident and the central claim of well-calibrated uncertainty collapses.

Editorial extensions

If this is right

  • Simulation-based inference from gravitational-wave catalogs can be run with emulator uncertainty marginalized over, using the posterior-predictive likelihood (the paper's Eq. 9) instead of conditioning on a single best-fit flow.
  • Data augmentation for rarely synthesized sub-populations, such as unequal-mass-ratio, high-chirp-mass binaries, high-metallicity systems, or large-natal-kick populations, comes with credible intervals that contain the truth, so feature predictions carry honest error bars without a full-scale simulator rerun.
  • Flow architectures can be ranked with the Bayesian information criterion computed from the sampled posterior, which penalizes over-complex models and reduces the epistemic uncertainty that comes from architecture choice.
  • For heavily downsampled training sets, the posterior-predictive emulator is closer to the true population than any individual maximum-likelihood retraining, so uncertainty marginalization is necessary for unbiased results.
  • The same Bayesian treatment transfers to continuous normalizing flows, whose far smaller parameter counts make HMC sampling tractable for higher-dimensional populations; the paper demonstrates the complexity reduction in its appendix.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The coverage claims inherit the choice of prior width, which the paper tunes on the held-out test set; a deployment-grade version of the method would need a hierarchical prior over the width or validation-only calibration, and its coverage on a genuinely new population is a direct prediction the paper does not yet test.
  • Calibration is reported as an average over parameter-space bins; the per-bin coverage in the sparsest tails, exactly the high-chirp-mass, unequal-mass-ratio region the data-augmentation use case targets, is the more demanding test and could in principle differ from the average.
  • The demonstration is limited to two initial conditions from an earlier generation of the simulator; carrying the method to the higher-dimensional, higher-fidelity simulation grids will plausibly require the continuous-flow variant the paper sketches, and the question is whether calibration survives that shift in dimensionality and compute budget.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper develops a Bayesian normalizing flow framework for emulating compact-binary population synthesis simulations. The flow is trained as a conditional density estimator of binary parameters given initial conditions, and uncertainty in the flow parameters is quantified by sampling the posterior with HMC/NUTS using the exact flow log-likelihood. The authors compute credible intervals for emulated densities and generated samples, compare calibration against an ensemble of MLE-trained flows, perform a BIC-based architecture comparison, and demonstrate data augmentation for a rarely populated BBH sub-population. The demonstration uses a POSYDON/COSMIC common-envelope BBH dataset, with a heavily downsampled training set (20 binaries per grid point) and a single held-out population (α=1.0, χb=0.1) as the test set.

Significance. If the calibration claim is robust, this is a timely and useful methodological contribution: it provides a concrete recipe for exact posterior sampling of normalizing-flow parameters, releases public code (the `naz` package), compares favorably against ensemble-based UQ on sparse training sets, and shows a practical data-augmentation application for a realistic PopSynth dataset. The use of the exact likelihood, rather than an approximate loss, is a genuine strength, as is the explicit comparison with MLE ensembles and the discussion of CNF-based scalability. However, the central claim of well-calibrated UQ is currently undermined by the way the prior width is tuned and evaluated on the same test set, and by the use of a data-dependent prior centered at the MLE. These issues are fixable but require changes to the experimental protocol and to the interpretation of the credible intervals.

major comments (4)
  1. [Sec. 6.2 and Sec. 9.5] The prior width σ0 is selected using the same held-out population (α=1.0, χb=0.1, Sec. 4) that is then used to report the calibration curves and the claim that the credible intervals 'indeed encompass the true population distribution the expected number of times.' Section 6.2 states that calibration statistics are used 'to optimize the choice of σ0,' and Sec. 9.5 reports that 'good calibration is observed for σ0 = 0.25' on that same test set. Tuning σ0 to maximize empirical coverage on the test set and then measuring coverage on the same test set introduces selection bias; the reported calibration is conditional on the test set and can appear well-calibrated even for an overconfident posterior. The authors should tune σ0 on a separate validation population (or via cross-validation on the training grid) and report calibration on a test population that is never used for any tuning decision.
  2. [Sec. 6.1 and Algorithm 1, line 1] The prior on flow parameters is centered at the MLE ω0 obtained from the same training data that appears in the likelihood of Eq. (6). Specifically, Algorithm 1 computes ω0 via Adam on the training data and then defines the prior as an affine-transformed uniform distribution supported within a factor σ0 of ω0. Consequently, the training data are used twice: once to center the prior and once in the likelihood. The resulting posterior is an empirical-Bayes / data-dependent-prior posterior, not a fully Bayesian posterior, and the credible intervals are conditional on data-derived prior information. This is load-bearing because the paper's central claim is that the intervals are well-calibrated Bayesian UQ. The authors should either (i) reframe the method as empirical Bayes and demonstrate calibration on data independent of the prior-centering step, (ii) use a prior whose parameters are fixed without reference to the training data, or (iii) explicitly quantify the sensitivity of the reported coverage to the prior centering.
  3. [Sec. 9 and Sec. 3] No convergence diagnostics are reported for the HMC/NUTS sampling, despite the claim in Sec. 3 that the sampling methods 'ensure stability and convergence.' The paper states in Sec. 9 that 4 chains with 500 tuning and 1500 sampling steps are used for a model with O(10^5–10^6) parameters (the default MAF has 16 layers with 150 hidden units each). For a posterior of this dimension, 1500 post-tuning samples per chain is very small, and without R-hat, bulk/tail ESS, trace plots, or a check of how many posterior samples lie near the prior boundaries, one cannot exclude non-convergence or prior-dominated sampling as the reason for the observed calibration. Please provide these diagnostics for representative flow parameters and report the fraction of posterior samples at or near the prior box edges, especially given the bounded uniform prior of Sec. 6.1.
  4. [Sec. 6.2] The calibration procedure splits the M posterior samples into m random subsets, generates 10^6 samples per posterior draw, and estimates empirical coverage as the fraction of m subsets (averaged over bins) for which the credible interval contains the test count. The choices of m, nQ (number of quantile bins), and the number of posterior draws per subset are not reported, and the coverage estimate is averaged over bins of a single test population, so the reported 'fraction of times' is not a variance-calibrated frequentist coverage over independent test sets. These details should be specified and the sensitivity of the calibration curves to m and nQ should be shown; otherwise the calibration curves in Fig. 5 are difficult to interpret quantitatively.
minor comments (4)
  1. [Throughout] There are numerous typographical errors, including 'futrue' (Abstract and Sec. 1), 'pannel' (Fig. 1 caption), 'intitial' (Sec. 2), 'tracktable' (Sec. 2), 'paralelizable' (Sec. 2), 'sp-spopulations' (Sec. 7), 'PoPSynth' (Sec. 3), and '⃗dttain' in Eq. (5). A careful proofreading pass is needed.
  2. [Sec. 9.4] The BIC model comparison is acknowledged by the authors to be dominated by the complexity term nω ln N, and the table shows a huge BIC gap between architectures whose log-likelihood difference is tiny relative to that term. In its current form the BIC comparison does not support a meaningful architecture selection; the abstract's mention of 'model comparison' should be tempered or replaced with a criterion (e.g., WAIC or cross-validated log-likelihood) that is not dominated by the parameter count.
  3. [Sec. 6.1 and Algorithm 1] The LogPrior implementation uses the elementwise condition (ω−ω0)/(σ×ω0) ∈ [−1,1], which is ill-defined if any component of ω0 is zero or close to zero. This is a practical detail that needs clarification, since the prior support is defined multiplicatively about ω0.
  4. [Sec. 7] Equation (9) proposes to marginalize over flow uncertainties in SBI by selecting posterior draws ω ∈ S_ω for each draw of λ. The procedure is described only at a high level; the paper should specify whether the same ω is reused across all λ draws in a chain and how the resulting posterior over λ is assembled. This is important for reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

The paper tunes the prior width σ0 by calibration on the held-out test set and then reports calibration on that same test set, so the central claim of well-calibrated UQ is partly self-confirming.

  1. fitted input called prediction [Sec. 6.1 and Algorithm 1 set the prior at the MLE; Sec. 9.5 reports the resulting calibration.]
    "We start by choosing a maximal scale ⃗ σ= {σ0}nω i=1, which amounts to assuming a priori that the highest likelihood model parameters are within a factor of at most σ0 of ⃗ ω0. ... As we demonstrate shortly, calibration statistics can be constructed for choosing optimal values of σ0. ... Algorithm 1: ω0 ← Adam(α, β1, β2, Loss, ωinit)."

    The prior center ω0 is obtained by maximizing the same likelihood, on the same training data, that is later used as the likelihood inside the posterior sampler. Thus the training data are used twice: once to produce the MLE that fixes the prior location, and again in LogLikelihood inside LogPost. The resulting 'posterior' is not a standard Bayesian posterior with a fixed prior; it is an empirical-Bayes posterior conditioned on a data-derived prior location, which can artificially concentrate the credible intervals. The paper's central calibration claim is evaluated on intervals generated from this data-dependent prior.

  2. fitted input called prediction [Sec. 6.2 uses the test set to choose σ0; Sec. 9.5 reports calibration on the same test set.]
    "To appraise the accuracy of the estimated uncertainties and optimize the choice of σ0, we construct metrics that compare the empirical coverage of a test set to the theoretical coverage predicted by Bayesian inference. ... good calibration is observed for σ0 = 0.25 ... In other words, the estimated credible intervals indeed encompass the true population distribution the expected number of times, when averaged over all regions of the parameter space, justifying this choice of σ0."

    The same held-out population (α = 1.0, χb = 0.1, Sec. 4) is used both to select σ0 by optimizing the calibration curve (Sec. 6.2) and to report the final calibration in Fig. 5 and Sec. 9.5. Because σ0 is explicitly chosen to make the empirical coverage match the nominal coverage on this test set, the reported 'well-calibrated' credible intervals are a selected result rather than an independent evaluation. The central claim that the Bayesian UQ is well-calibrated is therefore partly forced by construction, since the evaluation metric on the test set was the tuning objective for σ0.

full rationale

The core methodological contribution—sampling the exact normalizing-flow likelihood with HMC to obtain a posterior over flow parameters—is self-contained and not circular: the likelihood is the standard NF log-density, and the posterior sampling is a legitimate Monte Carlo procedure. There is no load-bearing self-citation chain or imported uniqueness theorem. However, the headline empirical claim of well-calibrated uncertainty quantification is weakened by two data-dependent choices. First, the prior is centered at the MLE ω0 computed from the same training set used in the posterior likelihood, so the data are used twice and the prior is not a fixed, independent object. Second, and more seriously, the prior width σ0 is selected by calibration on the held-out test set, and the same test set is then used to demonstrate that the credible intervals are well calibrated. This makes the reported coverage a tuned quantity, not an unbiased estimate of the method's true frequentist coverage. Because the abstract and Sec. 9.5 emphasize well-calibrated UQ as the central benefit of the method, the circularity score is elevated to 6; the remainder of the derivation does not reduce to its inputs.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a data-dependent prior centered at the MLE ω0 (Sec. 6.1), a test-set-tuned prior scale σ0 (Sec. 9.5), an i.i.d. sampling assumption, and an unverified NUTS convergence assumption. The flow model itself is not a new physical entity. The free parameters are the prior center and width, both derived from or tuned with the data.

free parameters (2)
  • prior scale σ0 = 0.25
    Width of the affine-transformed uniform prior around the MLE ω0; tuned via test-set calibration in Sec. 9.5 and directly controls the width of the posterior credible intervals.
  • MLE flow parameters ω0 = Adam-optimized on training data
    Center of the prior (Sec. 6.1); makes the prior data-dependent and the posterior conditional on a fitted point estimate.
assumptions (5)
  • domain assumption The normalizing flow model density p̂(θ|λ,ω) equals the true conditional density of the simulator output up to model error.
    Used in Eq. (2) and the likelihood in Eq. (5). Assumes the simulator distribution is smooth and the flow is sufficiently expressive.
  • standard math Training data points are i.i.d. samples from p(θ,λ)=p(θ|λ)pλ(λ).
    Justifies the expected KL divergence loss in Eq. (3).
  • ad hoc to paper The prior centered at the MLE ω0 with support within a factor σ0 is a valid representation of prior knowledge about flow parameters.
    Introduced in Sec. 6.1 to make HMC tractable; not derived from independent physics or prior elicitation, and it uses the training data twice.
  • ad hoc to paper NUTS converges to the posterior within 500 tuning and 1500 sampling steps for a model with O(10^5-10^6) parameters.
    No R-hat or ESS diagnostics are reported; convergence is only inferred from calibration curves on the test set.
  • domain assumption The single held-out grid point (α=1.0, χb=0.1) is representative for validating coverage of the emulator.
    Used as the test set in Sec. 4; does not test extrapolation to initial conditions outside the training grid.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Emulating compact binary population synthesis simulations with uncertainty quantification and model comparison using Bayesian normalizing flows." pith.science (2026). https://pith.science/paper/D577LF36

@misc{pith2026250605657,
  author       = {Pith},
  title        = {Pith review of: Emulating compact binary population synthesis simulations with uncertainty quantification and model comparison using Bayesian normalizing flows},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D577LF36}},
  note         = {Machine review of arXiv:2506.05657}
}
read the original abstract

Population synthesis simulations of compact binary coalescences~(CBCs) play a crucial role in extracting astrophysical insights from an ensemble of gravitational wave~(GW) observations. However, realistic simulations can be costly to implement for a dense grid of initial conditions. Normalizing flows can emulate population synthesis runs to enable simulation-based inference from observed catalogs and data augmentation for feature prediction in rarely synthesizable sub-populations. However, flow predictions can be wrought with uncertainties, especially for sparse training sets. In this work, we develop a method for quantifying and marginalizing uncertainties in the emulators by implementing the Bayesian Normalizing flow, a conditional density estimator constructed from Bayesian neural networks. Using the exact likelihood function naturally associated with density estimators, we sample the posterior distribution of flow parameters with suitably chosen priors to quantify and marginalize over flow uncertainties. We demonstrate the accuracy, calibration, inference, and data-augmentation impacts of the estimated uncertainties for simulations of binary black hole populations formed through common envelope evolution. We outline the applications of the proposed methodology in the context of simulation-based inference from growing GW catalogs and feature prediction, with state-of-the-art binary evolution simulators, now marginalized over model and data uncertainties.

Figures

Figures reproduced from arXiv: 2506.05657 by the authors.

Figure 1
Figure 1. Density estimation (left) for SBI, and data augmentation (right) for feature-prediction with a MAF-based PopSynth emulators. For the right pannel we focus on under-represented regions of parameter space. Each curve represents the effective spin distribution of BBHs in the test set that belong to an unequal mass ratio and high chirp mass sub-population. While S. Colloms et al. (2025) have extensively discussed SBI us… view at source ↗
Figure 2
Figure 2. Predictions of the Bayesian NF for prior scale σ0 = 0.25. Highest posterior density credible intervals on the marginal distribution (left), and the predictive joint distribution (right) for BBH chirp-mass and effective inspiral spin, are shown. it can be seen that the posterior predictive distribution is much closer to the truth than individual MLE reruns on different realizations of the thinned dataset, while also … view at source ↗
Figure 3
Figure 3. Uncertainty quantification for sample generation using Bayesian flows, for emulating the joint distribution of four BBH parameters For sample generation, it is straightforward to represent and marginalize higher-dimensional densities since the latter amounts to ignoring samples of the marginalized dimensions. Hence, for this demonstration, we emulate the four-dimensional distribution of BBH component masses, effecti… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Feature prediction with uncertainty quantification. 9.4. Model comparison To assess the efficiency of different flow architectures and avoid overfitting, we compute the BIC for our default architecture and compare it with that of a different (more complex) one, both fo…
Figure 5
Figure 5. Figure 5: Calibration curves for Bayesian flows (left and center) withvarious choices of the prior scale σ0 and an ensemble of flows trained using the MLE method (right). 10. DISCUSSION AND FUTURE PROSPECTS In this work, we developed a self-consistent method for quantifying epis…
Figure 6
Figure 6. Figure 6: Performance of a simple MAF(left) vs CNF(right), showing the emulated population (blue) on the test set compared with the true distribution (orange). rather approximating the intractable likelihood of observable parameters given high-dimensional data. If the dimen￾sion…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. High-dimensional reliability-oriented Shapley effect estimation with Normalizing Flows

    stat.ME 2026-06 unverdicted novelty 6.5 of 10

    Target Shapley effects for high-dimensional correlated reliability problems can be estimated from a single failing sample by rewriting closed target Sobol indices via conditional densities and fitting those densities ...

Reference graph

Works this paper leans on

79 extracted references · 13 canonical work pages · cited by 1 Pith paper

  1. [1]

    2015, Class

    Aasi, J., et al. 2015, Class. Quant. Grav., 32, 074001, doi: 10.1088/0264-9381/32/7/074001

    doi:10.1088/0264-9381/32/7/074001 2015

  2. [2]

    G., et al

    Abac, A. G., et al. 2024, Astrophys. J. Lett., 970, L34, doi: 10.3847/2041-8213/ad5beb

    doi:10.3847/2041-8213/ad5beb 2024

  3. [3]

    2020, Astrophys

    Abbott, R., et al. 2020, Astrophys. J. Lett., 896, L44, doi: 10.3847/2041-8213/ab960f

    doi:10.3847/2041-8213/ab960f 2020

  4. [4]

    2023, Phys

    Abbott, R., et al. 2023, Phys. Rev. X, 13, 011048, doi: 10.1103/PhysRevX.13.011048

    doi:10.1103/physrevx.13.011048 2023

  5. [5]

    2015, Class

    Acernese, F., et al. 2015, Class. Quant. Grav., 32, 024001, doi: 10.1088/0264-9381/32/2/024001

    doi:10.1088/0264-9381/32/2/024001 2015

  6. [6]

    2021, PTEP, 2021, 05A102, doi: 10.1093/ptep/ptab018

    Akutsu, T., et al. 2021, PTEP, 2021, 05A102, doi: 10.1093/ptep/ptab018

    doi:10.1093/ptep/ptab018 2021

  7. [7]

    J., Bavera, S

    Andrews, J. J., Bavera, S. S., Briel, M., et al. 2024, https://arxiv.org/abs/2411.02376

    arXiv 2024

  8. [8]

    Antonini, F., & Perets, H. B. 2012, Astrophys. J., 757, 27, doi: 10.1088/0004-637X/757/1/27 18

    doi:10.1088/0004-637x/757/1/27 2012

Show all 79 references
  1. [9]

    2023, https://arxiv.org/abs/2309.16314

    Arbel, J., Pitas, K., Vladimirova, M., & Fortuin, V. 2023, https://arxiv.org/abs/2309.16314

    2023 arXiv

  2. [10]

    2017, Astrophys

    Bartos, I., Kocsis, B., Haiman, Z., & M´ arka, S. 2017, Astrophys. J., 835, 165, doi: 10.3847/1538-4357/835/2/165

    2017 doi

  3. [11]

    S., Fragos, T., Qin, Y., et al

    Bavera, S. S., Fragos, T., Qin, Y., et al. 2020, Astronomy &; Astrophysics, 635, A97, doi: 10.1051/0004-6361/201936204

    2020 doi

  4. [12]

    2010, ApJL, 715, L138, doi: 10.1088/2041-8205/715/2/L138

    Belczynski, K., Dominik, M., Bulik, T., et al. 2010, ApJL, 715, L138, doi: 10.1088/2041-8205/715/2/L138

    2010 doi

  5. [13]

    E., et al

    Belczynski, K., Repetto, S., Holz, D. E., et al. 2016, ApJ, 819, 108, doi: 10.3847/0004-637X/819/2/108

    2016 doi

  6. [14]

    J., & Downing, J

    Benacquista, M. J., & Downing, J. M. B. 2013, Living Rev. Rel., 16, 4, doi: 10.12942/lrr-2013-4

    2013 doi

  7. [15]

    2023, https://arxiv.org/abs/2302.01312

    Berry, L., & Meger, D. 2023, https://arxiv.org/abs/2302.01312

    2023 arXiv

  8. [16]

    2018, https://arxiv.org/abs/1701.02434

    Betancourt, M. 2018, https://arxiv.org/abs/1701.02434

    2018 arXiv

  9. [17]

    2024, Mach

    Bieringer, S., Diefenbacher, S., Kasieczka, G., & Trabs, M. 2024, Mach. Learn. Sci. Tech., 5, 045044, doi: 10.1088/2632-2153/ad9136

    2024 doi

  10. [18]

    F., & Trabs, M

    Bieringer, S., Kasieczka, G., Steffen, M. F., & Trabs, M. 2023, https://arxiv.org/abs/2312.14027

    2023 arXiv

  11. [19]

    P., Jankowiak, M., et al

    Bingham, E., Chen, J. P., Jankowiak, M., et al. 2019, J. Mach. Learn. Res., 20, 28:1. http://jmlr.org/papers/v20/18-403.html

    2019

  12. [20]

    2019, doi: 10.3847/1538-4357/ab4a79

    Bouffanais, Y., Mapelli, M., Gerosa, D., et al. 2019, doi: 10.3847/1538-4357/ab4a79

    2019 doi

  13. [21]

    2021, Mon

    Bouffanais, Y., Mapelli, M., Santoliquido, F., et al. 2021, Mon. Not. Roy. Astron. Soc., 507, 5224, doi: 10.1093/mnras/stab2438

    2021 doi

  14. [22]

    2018,, 0.3.13 http://github.com/jax-ml/jax

    Bradbury, J., Frostig, R., Hawkins, P., et al. 2018,, 0.3.13 http://github.com/jax-ml/jax

    2018

  15. [23]

    2025, https://arxiv.org/abs/2502.03523

    Breivik, K. 2025, https://arxiv.org/abs/2502.03523

    2025 arXiv

  16. [24]

    2020, Astrophys

    Breivik, K., et al. 2020, Astrophys. J., 898, 71, doi: 10.3847/1538-4357/ab9d85

    2020 doi

  17. [25]

    A., & Farr, W

    Callister, T. A., & Farr, W. M. 2024, Phys. Rev. X, 14, 021005, doi: 10.1103/PhysRevX.14.021005

    2024 doi

  18. [26]

    2018, https://arxiv.org/abs/1709.01179

    Chen, C., Li, C., Chen, L., et al. 2018, https://arxiv.org/abs/1709.01179

    2018 arXiv

  19. [27]

    Q., Zevin, M., & Vitale, S

    Cheng, A. Q., Zevin, M., & Vitale, S. 2023, Astrophys. J., 955, 127, doi: 10.3847/1538-4357/aced98

    2023 doi

  20. [28]

    Colloms, S., Berry, C. P. L., Veitch, J., & Zevin, M. 2025, https://arxiv.org/abs/2503.03819

    2025 arXiv

  21. [29]

    R., Gair, J., et al

    Dax, M., Green, S. R., Gair, J., et al. 2021, Physical Review Letters, 127, doi: 10.1103/physrevlett.127.241103

    2021 doi

  22. [30]

    R., Gair, J., et al

    Dax, M., Green, S. R., Gair, J., et al. 2025, Nature, 639, 49, doi: 10.1038/s41586-025-08593-z

    2025 doi

  23. [31]

    2024, https://arxiv.org/abs/2408.15136

    Mishra-Sharma, S., & Louppe, G. 2024, https://arxiv.org/abs/2408.15136

    2024 arXiv

  24. [32]

    2023, https://arxiv.org/abs/2311.00377

    Dirmeier, S., Hong, Y., Xin, Y., & Perez-Cruz, F. 2023, https://arxiv.org/abs/2311.00377

    2023 arXiv

  25. [33]

    2019, https://arxiv.org/abs/1906.04032

    Durkan, C., Bekasov, A., Murray, I., & Papamakarios, G. 2019, https://arxiv.org/abs/1906.04032

    2019 arXiv

  26. [34]

    2023, Astrophys

    Edelman, B., Farr, B., & Doctor, Z. 2023, Astrophys. J., 946, 16, doi: 10.3847/1538-4357/acb5ed

    2023 doi

  27. [35]

    Andrews, J., Bavera, S

    Fragos, T., J. Andrews, J., Bavera, S. S., et al. 2023, The Astrophysical Journal Supplement Series, 264, 45, doi: 10.3847/1538-4365/ac90c1

    2023 doi

  28. [36]

    2016, https://arxiv.org/abs/1506.02142

    Gal, Y., & Ghahramani, Z. 2016, https://arxiv.org/abs/1506.02142

    2016 arXiv

  29. [37]

    2022, Engineering Applications of Artificial Intelligence, 115, 105151, doi: 10.1016/j.engappai.2022.105151

    Ganaie, M., Hu, M., Malik, A., Tanveer, M., & Suganthan, P. 2022, Engineering Applications of Artificial Intelligence, 115, 105151, doi: 10.1016/j.engappai.2022.105151

    2022

  30. [38]

    2021, Nature Astron., 5, 749, doi: 10.1038/s41550-021-01398-w

    Gerosa, D., & Fishbach, M. 2021, Nature Astron., 5, 749, doi: 10.1038/s41550-021-01398-w

    2021 doi

  31. [39]

    Grathwohl, W., Chen, R. T. Q., Bettencourt, J., Sutskever, I., & Duvenaud, D. 2018, https://arxiv.org/abs/1810.01367

    2018 arXiv

  32. [40]

    2025, https://arxiv.org/abs/2302.13425

    He, W., Jiang, Z., Xiao, T., Xu, Z., & Li, Y. 2025, https://arxiv.org/abs/2302.13425

    2025

  33. [41]

    2025, Phys

    Heinzel, J., Mould, M., ´Alvarez-L´ opez, S., & Vitale, S. 2025, Phys. Rev. D, 111, 063043, doi: 10.1103/PhysRevD.111.063043

    2025 doi

  34. [42]

    M., Wang, C., & Paisley, J

    Hoffman, M., Blei, D. M., Wang, C., & Paisley, J. 2013, https://arxiv.org/abs/1206.7051

    2013 arXiv

  35. [43]

    D., & Gelman, A

    Hoffman, M. D., & Gelman, A. 2011, https://arxiv.org/abs/1111.4246 H¨ ullermeier, E., & Waegeman, W. 2021, Machine Learning, 110, 457–506, doi: 10.1007/s10994-021-05946-3

    2011 arXiv

  36. [44]

    P., & Ba, J

    Kingma, D. P., & Ba, J. 2017, https://arxiv.org/abs/1412.6980

    2017 arXiv

  37. [45]

    2023, https://arxiv.org/abs/2210.02747

    Le, M. 2023, https://arxiv.org/abs/2210.02747

    2023 arXiv

  38. [46]

    Mandel, I., & de Mink, S. E. 2016, Mon. Not. Roy. Astron. Soc., 458, 2634, doi: 10.1093/mnras/stw379

    2016 doi

  39. [47]

    2022, Phys

    Mandel, I., & Farmer, A. 2022, Phys. Rept., 955, 1, doi: 10.1016/j.physrep.2022.01.003

    2022 doi

  40. [48]

    2020a, Proc

    Mapelli, M. 2020a, Proc. Int. Sch. Phys. Fermi, 200, 87, doi: 10.3254/ENFI200005

    doi

  41. [49]

    2020b, Front

    Mapelli, M. 2020b, Front. Astron. Space Sci., 7, 38, doi: 10.3389/fspas.2020.00038

    2020

  42. [50]

    M., & Moriya, T

    Marchant, P., Langer, N., Podsiadlowski, P., Tauris, T. M., & Moriya, T. J. 2016, Astron. Astrophys., 588, A50, doi: 10.1051/0004-6361/201628133

    2016 doi

  43. [51]

    2022, Mon

    Mastrogiovanni, S., Lamberts, A., Srinivasan, R., Bruel, T., & Christensen, N. 2022, Mon. Not. Roy. Astron. Soc., 517, 3432, doi: 10.1093/mnras/stac2850 19

    2022 doi

  44. [52]

    Mould, M., Gerosa, D., & Taylor, S. R. 2022, Phys. Rev. D, 106, 103013, doi: 10.1103/PhysRevD.106.103013

    2022 doi

  45. [53]

    E., & Vitale, S

    Mould, M., Wolfe, N. E., & Vitale, S. 2025, https://arxiv.org/abs/2504.07197

    2025 arXiv

  46. [54]

    2023, Mechanical Systems and Signal Processing, 205, 110796, doi: https://doi.org/10.1016/j.ymssp.2023.110796

    Nemani, V., Biggio, L., Huan, X., et al. 2023, Mechanical Systems and Signal Processing, 205, 110796, doi: https://doi.org/10.1016/j.ymssp.2023.110796

    2023

  47. [55]

    1976, in Structure and Evolution of Close Binary Systems, ed

    Paczynski, B. 1976, in Structure and Evolution of Close Binary Systems, ed. P. Eggleton, S. Mitton, & J. Whelan, Vol. 73, 75

    1976

  48. [56]

    J., Mohamed, S., & Lakshminarayanan, B

    Papamakarios, G., Nalisnick, E., Rezende, D. J., Mohamed, S., & Lakshminarayanan, B. 2021, https://arxiv.org/abs/1912.02762

    2021 arXiv

  49. [57]

    2018, https://arxiv.org/abs/1705.07057

    Papamakarios, G., Pavlakou, T., & Murray, I. 2018, https://arxiv.org/abs/1705.07057

    2018 arXiv

  50. [58]

    C., & Murray, I

    Papamakarios, G., Sterratt, D. C., & Murray, I. 2019, https://arxiv.org/abs/1805.07226

    2019 arXiv

  51. [59]

    2017, in NIPS-W

    Paszke, A., Gross, S., Chintala, S., et al. 2017, in NIPS-W

    2017

  52. [60]

    2011, ApJS, 192, 3, doi: 10.1088/0067-0049/192/1/3

    Paxton, B., Bildsten, L., Dotter, A., et al. 2011, ApJS, 192, 3, doi: 10.1088/0067-0049/192/1/3

    2011 doi

  53. [61]

    2013, ApJS, 208, 4, doi: 10.1088/0067-0049/208/1/4

    Paxton, B., Cantiello, M., Arras, P., et al. 2013, ApJS, 208, 4, doi: 10.1088/0067-0049/208/1/4

    2013 doi

  54. [62]

    2015, ApJS, 220, 15, doi: 10.1088/0067-0049/220/1/15

    Paxton, B., Marchant, P., Schwab, J., et al. 2015, ApJS, 220, 15, doi: 10.1088/0067-0049/220/1/15

    2015 doi

  55. [63]

    B., et al

    Paxton, B., Schwab, J., Bauer, E. B., et al. 2018, ApJS, 234, 34, doi: 10.3847/1538-4365/aaa5a8

    2018 doi

  56. [64]

    2019, ApJS, 243, 10, doi: 10.3847/1538-4365/ab2241

    Paxton, B., Smolec, R., Schwab, J., et al. 2019, ApJS, 243, 10, doi: 10.3847/1538-4365/ab2241

    2019 doi

  57. [65]

    2019, arXiv preprint arXiv:1912.11554

    Phan, D., Pradhan, N., & Jankowiak, M. 2019, arXiv preprint arXiv:1912.11554

    2019 arXiv

  58. [66]

    2025, Phys

    Plunkett, C., Mould, M., & Vitale, S. 2025, Phys. Rev. D, 112, 023039, doi: 10.1103/jv8h-4ggy

    2025 doi

  59. [67]

    2020, https://arxiv.org/abs/2009.09346

    Poli, M., Massaroli, S., Yamashita, A., Asama, H., & Park, J. 2020, https://arxiv.org/abs/2009.09346

    2020 arXiv

  60. [68]

    2006, Living Rev

    Postnov, K., & Yungelson, L. 2006, Living Rev. Rel., 9, 6, doi: 10.12942/lrr-2006-6

    2006 doi

  61. [69]

    2024, https://arxiv.org/abs/2404.03166

    Ray, A., Maga˜ na Hernandez, I., Breivik, K., & Creighton, J. 2024, https://arxiv.org/abs/2404.03166

    2024 arXiv

  62. [70]

    2023, Astrophys

    Ray, A., Maga˜ na Hernandez, I., Mohite, S., Creighton, J., & Kapadia, S. 2023, Astrophys. J., 957, 37, doi: 10.3847/1538-4357/acf452

    2023 doi

  63. [71]

    J., & Mohamed, S

    Rezende, D. J., & Mohamed, S. 2016, https://arxiv.org/abs/1505.05770

    2016 arXiv

  64. [72]

    2023, Astrophys

    Riley, J., & Mandel, I. 2023, Astrophys. J., 950, 80, doi: 10.3847/1538-4357/accf90

    2023 doi

  65. [73]

    2022, https://arxiv.org/abs/2211.09008

    Ruhe, D., Wong, K., Cranmer, M., & Forr´ e, P. 2022, https://arxiv.org/abs/2211.09008

    2022 arXiv

  66. [74]

    2018, Astrophys

    Talbot, C., & Thrane, E. 2018, Astrophys. J., 856, 173, doi: 10.3847/1538-4357/aab34c van den Heuvel, E. P. J., Portegies Zwart, S. F., & de Mink, S. E. 2017, Mon. Not. Roy. Astron. Soc., 471, 4256, doi: 10.1093/mnras/stx1430 van Son, L. A. C., de Mink, S. E., Callister, T., e...

    2018 doi

  67. [75]

    2003, Astrophys

    Wen, L. 2003, Astrophys. J., 598, 419, doi: 10.1086/378794

    2003 doi

  68. [76]

    M., & Dimakis, A

    Whang, J., Lindgren, E. M., & Dimakis, A. G. 2021, https://arxiv.org/abs/2002.11743

    2021 arXiv

  69. [77]

    Wong, K. W. K., Breivik, K., Kremer, K., & Callister, T. 2021, Phys. Rev. D, 103, 083021, doi: 10.1103/PhysRevD.103.083021

    2021 doi

  70. [78]

    Zevin, M., Spera, M., Berry, C. P. L., & Kalogera, V. 2020, Astrophys. J. Lett., 899, L1, doi: 10.3847/2041-8213/aba74e

    2020 doi

  71. [79]

    S., Berry, C

    Zevin, M., Bavera, S. S., Berry, C. P. L., et al. 2021, Astrophys. J., 910, 152, doi: 10.3847/1538-4357/abe40e

    2021 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.