pith. sign in

arxiv: 2606.10762 · v1 · pith:HTKBEX5Pnew · submitted 2026-06-09 · 🌌 astro-ph.SR · astro-ph.IM

Amortized Simulation-Based Inference of Colliding-Wind Binaries from Short, Noisy Image Time Series

Niklas Kn\"oll , Tobias Buck , Lorenzo Branca , Giuseppe Viterbo This is my paper

Pith reviewed 2026-06-27 11:57 UTC · model grok-4.3

classification 🌌 astro-ph.SR astro-ph.IM
keywords colliding-wind binariessimulation-based inferenceamortized inferencestellar windsH-alpha imagingneural spline flowsparameter inference
0
0 comments X

The pith

A factorized spatio-temporal neural architecture infers seven physical parameters of colliding-wind binaries from short noisy H-alpha image sequences.

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

The paper establishes a complete amortized simulation-based inference pipeline for colliding-wind binaries. It processes 10-frame photon-count time series by first encoding local spatial morphology and then aggregating temporal information across frames. This separation produces time-translation equivariant representations that support reliable posterior inference even when photon counts are low. The resulting posteriors for mass-loss rates, wind velocities, and orbital elements are shown to be well-calibrated on synthetic data with realistic detector noise.

Core claim

A factorized spatio-temporal architecture that separates spatial encoding from temporal aggregation, when paired with a neural spline flow, enables amortized posterior inference of seven physical parameters from synthetic 10-frame H-alpha photon-count time series of colliding-wind binaries, with the posteriors verified as well-calibrated via TARP and SBC diagnostics.

What carries the argument

The factorized spatio-temporal architecture that separates spatial encoding from temporal aggregation and induces time-translation equivariance in the learned representation.

Load-bearing premise

The hydrodynamic simulations used to generate the training data accurately represent the true emission and morphology of colliding-wind binaries under the observed conditions.

What would settle it

Running the trained model on new synthetic time series generated from hydrodynamic simulations that deliberately alter wind collision morphology or emission physics and checking whether the recovered posteriors remain accurate and calibrated.

Figures

Figures reproduced from arXiv: 2606.10762 by Giuseppe Viterbo, Lorenzo Branca, Niklas Kn\"oll, Tobias Buck.

Figure 1
Figure 1. Figure 1: SBI flow chart for the CWB problem. From a prior p(θ) over seven physical parameters, we run 3D hydrodynamic simulations with a modified version of the JAX-based fluid solver astronomix (Storcks, 2025) together with an orbit integrator and an Hα emissivity + Poisson noise pipeline to produce 10-frame photon-count time series. The resulting {(θi, xi)} pairs train a neural spline flow with a spatio-temporal … view at source ↗
Figure 2
Figure 2. Figure 2: Forward-model pipeline for a representative CWB (N = 2563 ; x and y in grid units). Left to right: mass-density slice in the orbital plane; Hα emissivity jHα (orbital plane slice); and line-of-sight-integrated Hα intensity J. The intensity traces the high-density bow-shock apex and the wind-collision region bounded by the two contact discontinuities. 4. Simulation-Based Inference We target the posterior p(… view at source ↗
Figure 3
Figure 3. Figure 3: True vs. inferred values for a N = 200 subset of the test simulations. Error bars are 1σ intervals estimated from 2000 posterior samples. Mass-loss rates are recovered accurately over the 4-decade range, while terminal wind velocities slightly compress to the mean. The eccentricity e and the inclination i show phase-dependent recoverability. The turbulence parameter η remains at the prior. gregator, and tw… view at source ↗
Figure 4
Figure 4. Figure 4: Inferred posterior distributions and parameter correlations for two reference CWBs. Diagonal panels show the one-dimensional marginal posteriors with the shaded band marking the central 68% credible interval and the dashed line the posterior median. Off-diagonal panels show the two-dimensional joint posteriors as filled contours enclosing the 0.5σ, 1σ, 1.5σ and 2σ credible regions. Solid red lines indicate… view at source ↗
Figure 5
Figure 5. Figure 5: Complete 10-frame Hα photon-count time series for Case 2 (low wind luminosity, high relative noise). Noise dominates in most snapshots and shock features and the stagnation point are only intermittently visible. However, the terminal-velocity posteriors approximately coincide with the prior, because the distinctive features that encode v∞ (opening angle, width of the shocked region, and periastron asymmetr… view at source ↗
Figure 6
Figure 6. Figure 6: shows the empirical mass / mass-loss relation M0(M˙ ) used to assign a stellar mass to each sampled M˙ (with 5% Gaussian scatter), together with the prior-induced distribution of orbital periods for a binary with semi-major axis a = 10 AU. (a) M0(M˙ ) with 5% Gaussian scatter. (b) Prior-induced period distribution [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Mass-density slice in the orbital plane of a stationary CWB, showing the stand-off distance R0 measured from the stronger￾wind star A (Eq. 8) and the opening angle θ of the bow-shock cone (becomes θ∞ for r → ∞; Eq. 9). Example parameters: M˙ 1 = 5.2 · 10−6 M⊙/yr, M˙ 2 = 5.2 · 10−7 M⊙/yr, v∞,1 = 2200 km/s, v∞,2 = 2000 km/s, a = 5 AU C. Hyperparameter tuning [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: shows the 128-trial Pareto front optimization plot, in which the validation NLL and the mean TARP deviation from the ideal coverage curve are optimized simultaneously. The green marked and annotated trial 20 giving the best final calibration among the chosen best trials for a training on the complete test simulation set (N = 40, 000) [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: RMSE and posterior width as a function of training-set size. Both metrics saturate slowly past N ≈ 2 · 104 simulations. E. Calibration diagnostics [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: shows the joint TARP coverage curve and the per-parameter SBC plot for the trained model. The TARP curve tracks the ideal diagonal across all credibility levels, indicating that the full 7-dimensional posterior is jointly well-calibrated (ATC = −0.061). The SBC as an empirical CDF vs. posterior rank plot confirms an approximately uniform rank distribution for mass-loss rate 1, both terminal wind velocitie… view at source ↗
Figure 11
Figure 11. Figure 11: shows the full 10-frame time series for Case 1 of [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: shows the full posterior corner plot for Case 3 of [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: ABC-rejection posteriors obtained from the 62-D hand-crafted summary of Eq. 10 for the reference cases 1 and 2 (see Tab. 3). Only the mass-loss rates are constrained, with widths noticeably broader (1σ CI spans around a full decade for M˙ 1/2 in case 1) than those of the neural estimator. All other posterior distributions closely resemble their respective prior distributions. Structural limitations of ABC… view at source ↗
read the original abstract

Colliding-wind binaries (CWBs), which are systems of two massive stars whose supersonic winds collide into bow shocks, encode rich information about stellar wind properties in their multi-frequency emission, e.g. images in the H$\alpha$, X-ray, and radio wavelengths. Inferring physical parameters (mass-loss rates, terminal wind velocities, orbital elements) from short time-series observations is a compelling but challenging inverse problem, because the forward hydrodynamic simulator is computationally expensive and the likelihood is intractable. We adopt a factorized spatio-temporal architecture for amortized posterior inference that separates spatial encoding from temporal aggregation. This design aligns with the structure of the underlying physical process of local morphology and global dynamical evolution, induces time-translation equivariance in the learned representation, and improves identifiability in low-signal regimes. Coupled with a neural spline flow conditioned on these spatio-temporal embeddings of 10-frame H$\alpha$ photon-count time series, we present a complete simulation-based inference pipeline for CWBs. Our method jointly infers seven physical parameters from synthetic observations under realistic detector noise, with posteriors verified as well-calibrated via TARP and SBC diagnostics. The approach naturally expands posterior width in information-poor regimes (low photon counts) and robustly recovers orbital parameters and mass-loss rates, demonstrating the feasibility of amortized likelihood-free inference for this challenging astrophysical inverse problem.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 2 minor

Summary. The manuscript develops an amortized simulation-based inference pipeline for colliding-wind binaries that uses a factorized spatio-temporal neural architecture (spatial encoding followed by temporal aggregation) to produce embeddings for a neural spline flow. The method jointly infers seven physical parameters from synthetic 10-frame Hα photon-count time series generated by hydrodynamic simulations plus realistic detector noise. Posteriors are reported as well-calibrated on held-out synthetic data according to TARP and SBC diagnostics, with the architecture chosen to induce time-translation equivariance and to widen posteriors in low-signal regimes.

Significance. If the reported calibration on synthetic data holds, the work establishes a feasible amortized SBI approach for an inverse problem whose forward model is computationally expensive and whose likelihood is intractable. The explicit alignment of the network architecture with the separation of local morphology and global orbital evolution is a methodological strength, as is the demonstration that posterior width naturally expands under low photon counts. Because the central claim is confined to performance on data drawn from the same simulation distribution used for training, the result is a solid proof-of-concept for likelihood-free inference in this domain rather than a direct constraint on real observations.

minor comments (2)
  1. [Methods] The abstract and methods description refer to 'realistic detector noise' without an explicit equation or table specifying the noise model (e.g., Poisson statistics, read noise, or PSF convolution); adding this detail would improve reproducibility.
  2. [Results] The claim that the architecture 'improves identifiability in low-signal regimes' would benefit from a quantitative comparison (e.g., a table of posterior widths or coverage probabilities) against a non-factorized baseline in the results section.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their accurate summary of the manuscript and for recommending minor revision. The assessment correctly identifies the work as a proof-of-concept demonstration of amortized SBI on synthetic data drawn from the training distribution.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper describes an amortized SBI pipeline that trains a spatio-temporal encoder plus neural spline flow on synthetic Hα time series generated from external hydrodynamic simulations, then evaluates posterior calibration on held-out draws from the same simulator using TARP and SBC. The central claim is restricted to recovery of seven parameters under the model's own noise model; no derivation step equates a fitted quantity to a prediction by construction, no uniqueness theorem is imported from self-citation, and no ansatz is smuggled via prior work. The pipeline is therefore self-contained against its stated synthetic benchmark.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the fidelity of the hydrodynamic forward model used to create training data; no new physical entities are introduced.

axioms (1)
  • domain assumption Hydrodynamic simulations of colliding winds accurately capture the relevant physics for generating training data
    The entire inference pipeline depends on the quality of these simulations as the forward model.

pith-pipeline@v0.9.1-grok · 5785 in / 1149 out tokens · 29555 ms · 2026-06-27T11:57:44.912280+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 27 canonical work pages · 1 internal anchor

  1. [1]

    arXiv e-prints , keywords =

    A COMPASS to Model Comparison and Simulation-Based Inference in Galactic Chemical Evolution. arXiv e-prints , keywords =. doi:10.48550/arXiv.2507.05060 , archivePrefix =. 2507.05060 , primaryClass =

    work page doi:10.48550/arxiv.2507.05060

  2. [2]

    FlowSN: Neural Simulation-Based Inference under Realistic Selection Effects applied to Supernova Cosmology

    FlowSN: Neural Simulation-Based Inference under Realistic Selection Effects applied to Supernova Cosmology. arXiv e-prints , keywords =. doi:10.48550/arXiv.2603.11165 , archivePrefix =. 2603.11165 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2603.11165

  3. [3]

    Proceedings of the National Academy of Sciences , volume =

    The frontier of simulation-based inference. Proceedings of the National Academy of Science , keywords =. doi:10.1073/pnas.1912789117 , archivePrefix =. 1911.01429 , primaryClass =

    work page doi:10.1073/pnas.1912789117 1911

  4. [4]

    arXiv e-prints , keywords =

    Some models are useful, but how do we know which ones? Towards a unified Bayesian model taxonomy. arXiv e-prints , keywords =. doi:10.48550/arXiv.2209.02439 , archivePrefix =. 2209.02439 , primaryClass =

    work page doi:10.48550/arxiv.2209.02439

  5. [5]

    arXiv e-prints , keywords =

    JANA: Jointly Amortized Neural Approximation of Complex Bayesian Models. arXiv e-prints , keywords =. doi:10.48550/arXiv.2302.09125 , archivePrefix =. 2302.09125 , primaryClass =

    work page doi:10.48550/arxiv.2302.09125

  6. [6]

    arXiv e-prints , keywords =

    Emulating Radiative Transfer in Astrophysical Environments. arXiv e-prints , keywords =. doi:10.48550/arXiv.2511.08219 , archivePrefix =. 2511.08219 , primaryClass =

    work page doi:10.48550/arxiv.2511.08219

  7. [7]

    Machine Learning: Science and Technology , keywords =

    Robust simulation-based inference in cosmology with Bayesian neural networks. Machine Learning: Science and Technology , keywords =. 2023a , month = mar, volume =. doi:10.1088/2632-2153/acbb53 , archivePrefix =. 2207.08435 , primaryClass =

    work page doi:10.1088/2632-2153/acbb53

  8. [8]

    and Macke, Jakob H

    Flow Matching for Scalable Simulation-Based Inference. Machine Learning for Astrophysics , year = 2023, month = jul, eid =. doi:10.48550/arXiv.2305.17161 , archivePrefix =. 2305.17161 , primaryClass =

    work page doi:10.48550/arxiv.2305.17161 2023

  9. [9]

    arXiv e-prints , keywords =

    CASBI -- Chemical Abundance Simulation-Based Inference for Galactic Archeology. arXiv e-prints , keywords =. doi:10.48550/arXiv.2411.17269 , archivePrefix =. 2411.17269 , primaryClass =

    work page doi:10.48550/arxiv.2411.17269

  10. [10]

    , keywords =

    Inferring Galactic parameters from chemical abundances with simulation-based inference. , keywords =. doi:10.1051/0004-6361/202554306 , archivePrefix =. 2503.02456 , primaryClass =

    work page doi:10.1051/0004-6361/202554306

  11. [11]

    , keywords =

    The dynamical memory of tidal stellar streams: Joint inference of the Galactic potential and the progenitor of GD-1 with flow matching. , keywords =. doi:10.1051/0004-6361/202558358 , archivePrefix =. 2512.04600 , primaryClass =

    work page doi:10.1051/0004-6361/202558358

  12. [12]

    2025 , copyright =

    Storcks, Leonard , title =. 2025 , copyright =. doi:10.5281/ZENODO.17782162 , url =

    work page doi:10.5281/zenodo.17782162 2025

  13. [13]

    , keywords =

    Exact, Algebraic Solutions of the Thin-Shell Two-Wind Interaction Problem. , keywords =. doi:10.1086/177820 , adsurl =

    work page doi:10.1086/177820

  14. [14]

    Monthly Notices of the Royal Astronomical Society , volume =

    Pejcha, Ondřej and Calderón, Diego and Kurfürst, Petr , title =. Monthly Notices of the Royal Astronomical Society , volume =. 2022 , month =. doi:10.1093/mnras/stab3729 , url =

    work page doi:10.1093/mnras/stab3729 2022

  15. [15]

    2023b , eprint=

    Sampling-Based Accuracy Testing of Posterior Estimators for General Inference , author=. 2023b , eprint=

  16. [16]

    arXiv preprint arXiv:1804.06788 , year=

    Validating Bayesian inference algorithms with simulation-based calibration , author=. arXiv preprint arXiv:1804.06788 , year=

    arXiv

  17. [17]

    Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics , pages =

    Sequential Neural Likelihood: Fast Likelihood-free Inference with Autoregressive Flows , author =. Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics , pages =. 2019 , editor =

    2019

  18. [18]

    Monthly Notices of the Royal Astronomical Society , volume =

    Kochanek, C S , title =. Monthly Notices of the Royal Astronomical Society , volume =. 2019 , month =. doi:10.1093/mnras/sty3363 , url =

    work page doi:10.1093/mnras/sty3363 2019

  19. [19]

    2023 , publisher =

    Taming rotationally supported disks using state of the art numerical methods , author =. 2023 , publisher =

    2023

  20. [20]

    2024 , eprint=

    Differentiable Conservative Radially Symmetric Fluid Simulations and Stellar Winds -- jf1uids , author=. 2024 , eprint=

    2024

  21. [21]

    Townsend, R. H. D. , year=. AN EXACT INTEGRATION SCHEME FOR RADIATIVE COOLING IN HYDRODYNAMICAL SIMULATIONS , volume=. The Astrophysical Journal Supplement Series , publisher=. doi:10.1088/0067-0049/181/2/391 , number=

    work page doi:10.1088/0067-0049/181/2/391

  22. [22]

    2025 , eprint=

    The Shamrock code: I- Smoothed Particle Hydrodynamics on GPUs , author=. 2025 , eprint=

    2025

  23. [23]

    and Ashkar, H

    Aharonian, F. and Ashkar, H. and Backes, M. and Barbosa Martins, V. and Becherini, Y. and Berge, D. and Bi, B. and Böttcher, M. and de Bony de Lavergne, M. and Bradascio, F. and Brose, R. and Brun, F. and Bulik, T. and Burger-Scheidlin, C. and Cangemi, F. and Caroff, S. and Casanova, S. and Cerruti, M. and Chand, T. and Chandra, S. and Chen, A. and Chibue...

    work page doi:10.1051/0004-6361/202244323

  24. [24]

    2025 , eprint=

    Differentiable Cosmological Hydrodynamics for Field-Level Inference and High Dimensional Parameter Constraints , author=. 2025 , eprint=

    2025

  25. [25]

    10.1051/0004-6361/201834832

    Thermal emission from bow shocks - I. 2D hydrodynamic models of the Bubble Nebula , DOI= "10.1051/0004-6361/201834832", url= "https://doi.org/10.1051/0004-6361/201834832", journal =

    work page doi:10.1051/0004-6361/201834832

  26. [26]

    10.1051/0004-6361/202243531

    Thermal emission from bow shocks - II. 3D magnetohydrodynamic models of ζ Ophiuchi , DOI= "10.1051/0004-6361/202243531", url= "https://doi.org/10.1051/0004-6361/202243531", journal =

    work page doi:10.1051/0004-6361/202243531

  27. [27]

    10.1051/0004-6361/202553722

    Thermal emission from bow shocks - III. Variable diffuse X-ray emission from stellar-wind bow shocks driven by dynamical instabilities , DOI= "10.1051/0004-6361/202553722", url= "https://doi.org/10.1051/0004-6361/202553722", journal =

    work page doi:10.1051/0004-6361/202553722

  28. [28]

    Jan Boelts and Michael Deistler and Manuel Gloeckler and Álvaro Tejero-Cantero and Jan-Matthis Lueckmann and Guy Moss and Peter Steinbach and Thomas Moreau and Fabio Muratore and Julia Linhart and Conor Durkan and Julius Vetter and Benjamin Kurt Miller and Maternus Herold and Abolfazl Ziaeemehr and Matthijs Pals and Theo Gruner and Sebastian Bischoff and ...

    work page doi:10.21105/joss.07754 2025

  29. [29]

    RAS Techniques and Instruments , volume =

    Spurio Mancini, A and Docherty, M M and Price, M A and McEwen, J D , title =. RAS Techniques and Instruments , volume =. 2023 , month =. doi:10.1093/rasti/rzad051 , url =

    work page doi:10.1093/rasti/rzad051 2023

  30. [30]

    2018 , eprint=

    Masked Autoregressive Flow for Density Estimation , author=. 2018 , eprint=

    2018

  31. [31]

    2018 , eprint=

    Fast -free Inference of Simulation Models with Bayesian Conditional Density Estimation , author=. 2018 , eprint=

    2018

  32. [32]

    Proceedings of the 36th International Conference on Machine Learning , pages =

    Automatic Posterior Transformation for Likelihood-Free Inference , author =. Proceedings of the 36th International Conference on Machine Learning , pages =. 2019 , editor =

    2019

  33. [33]

    Proceedings of the 25th

    Optuna: A Next-generation Hyperparameter Optimization Framework , author=. Proceedings of the 25th

  34. [34]

    and Pratap, A

    Deb, K. and Pratap, A. and Agarwal, S. and Meyarivan, T. , journal=. A fast and elitist multiobjective genetic algorithm: NSGA-II , year=

  35. [35]

    2019 , eprint=

    Neural Spline Flows , author=. 2019 , eprint=

    2019

  36. [36]

    Astrophysics of gaseous nebulae and active galactic nuclei

  37. [37]

    HOW-MHD: A High-order WENO-based Magnetohydrodynamic Code with a High-order Constrained Transport Algorithm for Astrophysical Applications , volume=

    Seo, Jeongbhin and Ryu, Dongsu , year=. HOW-MHD: A High-order WENO-based Magnetohydrodynamic Code with a High-order Constrained Transport Algorithm for Astrophysical Applications , volume=. The Astrophysical Journal , publisher=. doi:10.3847/1538-4357/acdf4b , number=

    work page doi:10.3847/1538-4357/acdf4b

  38. [38]

    10.1051/0004-6361/201117751

    Grids of stellar models with rotation - I. Models from 0.8 to 120 M solar metallicity (Z = 0.014) , DOI= "10.1051/0004-6361/201117751", url= "https://doi.org/10.1051/0004-6361/201117751", journal =

    work page doi:10.1051/0004-6361/201117751

  39. [39]

    The relative impact of photoionizing radiation and stellar winds on different environments , volume=

    Haid, S and Walch, S and Seifried, D and Wünsch, R and Dinnbier, F and Naab, T , year=. The relative impact of photoionizing radiation and stellar winds on different environments , volume=. Monthly Notices of the Royal Astronomical Society , publisher=. doi:10.1093/mnras/sty1315 , number=

    work page doi:10.1093/mnras/sty1315

  40. [40]

    Genetics , volume =

    Beaumont, Mark A and Zhang, Wenyang and Balding, David J , title =. Genetics , volume =. 2002 , month =. doi:10.1093/genetics/162.4.2025 , url =

    work page doi:10.1093/genetics/162.4.2025 2002

  41. [41]

    A Closer Look at Spatiotemporal Convolutions for Action Recognition , year=

    Tran, Du and Wang, Heng and Torresani, Lorenzo and Ray, Jamie and LeCun, Yann and Paluri, Manohar , booktitle=. A Closer Look at Spatiotemporal Convolutions for Action Recognition , year=

  42. [42]

    Long-term recurrent convolutional networks for visual recognition and description , year=

    Donahue, Jeff and Hendricks, Lisa Anne and Guadarrama, Sergio and Rohrbach, Marcus and Venugopalan, Subhashini and Darrell, Trevor and Saenko, Kate , booktitle=. Long-term recurrent convolutional networks for visual recognition and description , year=

  43. [43]

    S., & Fabian, A

    Damineli, A. and Hillier, D. J. and Corcoran, M. F. and Stahl, O. and Levenhagen, R. S. and Leister, N. V. and Groh, J. H. and Teodoro, M. and Colombo, J. F. Albacete and Gonzalez, F. and Arias, J. and Levato, H. and Grosso, M. and Morrell, N. and Gamen, R. and Wallerstein, G. and Niemela, V. , year=. The periodicity of the Carinae events , volume=. Month...

    work page doi:10.1111/j.1365-2966.2007.12815.x 2007