Emulating compact binary population synthesis simulations with robust uncertainty quantification and model comparison: Bayesian normalizing flows. Pytorch, Pyro, JAX, Numpyro

Anarya Ray · 2025

Read Emulating compact binary population synthesis simulations with robust uncertainty quantification and model comparison: Bayesian normalizing flows. Pytorch, Pyro, JAX, Numpyro by Anarya Ray (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-05657.txt.

Introduction

Anarya Ray builds a fast stand-in for costly simulations of merging black holes and neutron stars, in this 2025 paper. Population synthesis codes predict how often such pairs form and merge, feeding the reading of gravitational wave catalogs. Running them across a dense grid of starting conditions is expensive. A learned emulator could fill the grid cheaply, but only if its own doubts are measured honestly.

Ray trains Bayesian normalizing flows, flexible density models, to mimic the simulator's output. The Bayesian twist puts error bars on the emulator itself, which can then be folded into later analysis. That matters most where training runs are sparse and the flow must guess. The method also supports comparing rival formation models against real wave data. The paper was written with a modern Python stack for such models.

Catalog record

CREATORAnarya Ray
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE53 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-05657

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Emulating compact binary population synthesis simulations with robust uncertainty quantification and model comparison: Bayesian normalizing flows. Pytorch, Pyro, JAX, Numpyro Anarya Ray 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 the distribution functions of a simulated population of binary parameters and thereby enable empirical constraints on the astrophysical initial conditions and branching fractions of various formation channels, given data from a catalog of GW observations. They can also be used for data amplification in sparse regions of the CBC parameter space to better guide the development of phenomenological population models for rarely synthesizable systems with components in theorized mass gaps, without having to simulate a prohibitively large number of binaries. However, flow predictions are wrought with uncertainties, especially for sparse training sets. In this work, I develop a method for quantifying and marginalizing uncertainties in the emulators by introducing the Bayesian Normalizing flow, a conditional density estimator constructed from Bayesian neural networks. Using the exact likelihood function naturally associated with density estimators, I sample the posterior distribution of flow parameters with suitably chosen priors to quantify and marginalize over flow uncertainties. I demonstrate the accuracy, calibration, and data-amplification impacts of the estimated uncertainties for simulations of binary black hole populations formed through common envelope evolution. I 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. I also sketch other applications for general simulation-based approaches in GW astronomy. show][email protected] I Introduction The origin of merging compact binaries (CBCs) observed by ground-based gravitational wave (GW) detectors ( J. Aasi et al. 2015 ; F. Acernese et al. 2015 ; T. Akutsu et al. 2021 , such as the LVK, ) remains an open question ( I. Mandel & A. Farmer 2022 ; M. Mapelli 2020a ; M. Mapelli 2020b ) . Several formation scenarios have been proposed in existing literature, most of which are characterized by partially unconstrained physics. Broadly, CBCs capable of merging within a Hubble time are expected to emerge from isolated evolution of massive stellar binaries that undergo orbital hardening, via either common-envelope, stable mass transfer, or chemical mixing ( K. Postnov & L. Yungelson 2006 ; B. Paczynski 1976 ; E. P. J. van den Heuvel et al. 2017 ; P. Marchant et al. 2016 ; I. Mandel & S. E. de Mink 2016 , e.g., ) , dynamical assembly assisted by either a tertiary companion, multiple exchanges in dense clusters, or gas-assisted migration ( L. Wen 2003 ; F. Antonini & H. B. Perets 2012 ; M. J. Benacquista & J. M. B. Downing 2013 ; I. Bartos et al. 2017 , e.g.,) , and hierarchical mergers in star clusters and AGN disks ( D. Gerosa & M. Fishbach 2021 ) . The uncertain initial conditions and parameters that underlie these evolutionary pathways are crucial to elucidating unconstrained astrophysical phenomena and are often imprinted on the measurable population properties of CBC observations. Population synthesis ( K. Breivik 2025 , PopSynth, ) simulations can model these astrophysical imprints on the various features and correlations of the population-level distributions of measurable CBC parameters. PopSynth predictions can guide the construction of targeted phenomenological models ( C. Talbot & E. Thrane 2018 ; L. A. C. van Son et al. 2022 , e.g.,) of the distribution function whose parameters can be constrained given GW data from growing catalogs to extract astrophysical insights ( R. Abbott et al. 2023a ) . More flexible population models ( B. Edelman et al. 2023 ; T. A. Callister & W. M. Farr 2024 ; A. Ray et al. 2023 ; A. Ray et al. 2024 ; J. Heinzel et al. 2025 , e.g.,) are also employed to search for additional physics in the data beyond the assumptions of astrophysically motivated density functions, with the reconstructed population features being harder to interpret in terms of the physical processes underlying an evolutionary channel. On the other hand, direct comparison of PopSynth outputs to a collection of measured CBC parameters by means of density estimation can provide observational constraints on astrophysical initial conditions of the simulated evolutionary pathways ( M. Zevin et al. 2021 ; J. Riley & I. Mandel 2023 ; S. Colloms et al. 2025 ) . Such approaches can also be considered reasonably flexible in terms of the underlying physical scenarios explored provided a diverse collection of simulated formation channels are used in the inference. Normalizing flows ( D. J. Rezende & S. Mohamed 2016 ; G. Papamakarios et al. 2021 , NFs,) are powerful generative models that can assist with both simulation-based inference (SBI) and feature prediction, from PopSynth outputs. For example, S. Colloms et al. 2025 use NFs for estimating the conditional distribution of CBC parameters for each simulated population given values of the astrophysical initial conditions that characterize the simulated pathways. Once trained, a weighted mixture of different estimators, each corresponding to a specific formation channel, can facilitate tractable interpolation of the population likelihood across discrete gridpoints of simulation initial conditions, which can be constrained simultaneously with the branching fractions (mixture weights) of each channel given GW data from growing catalogs. S. Colloms et al. 2025 constrain the common envelop ejection effeciency, an uncertain parameter in isolated binary stellar evolution, the natal spin of black holes (BHs), and the branching fractions of all five formation channels whose simulations are used in their inference, from the confident binary black hole (BBH) detections (excluding population outliers) reported in the third GW transient catalog ( R. Abbott et al. 2023b , GWTC-3) . On the other hand, NF emulators can also be used for data amplification to predict population-level features and correlations in underrepresented regions of the CBC parameter space, where very few systems are synthesized from a particular channel, without having to simulate a prohibitively large ensemble of binaries. However, like other deep learning emulators, the predictions of normalizing flows are susceptible to epistemic (model) and aleatoric (data) uncertainties ( V. Nemani et al. 2023 ; W. He et al. 2025 ) , which can, in principle, bias the inferred astrophysical conclusions. For flow-based emulators, model uncertainties may arise either from inefficiencies in parameter learning, sub-optimal architectures, and out of distribution samples in the observed data, where as data uncertainties may result from the inherent sparsity and randomness of the training set which may be caused by particular formation channels being highly inefficient in certain regions of the CBC parameter space or by numerical approximation used by the simulators internally to speed up computation ( V. Nemani et al. 2023 ; W. He et al. 2025 ; J. J. Andrews et al. 2024 ) . While biases associated with ignoring model uncertainties may, in principle, be reduced by generating more training data, costly simulators can prohibit such endeavors. On the other hand, astrophysical bottlenecks leading to sparse and noisy datasets cannot be mitigated by simulating a larger number of binaries. In other words, for state

Read the complete text