Rapid and robust simulation-based inference for kilonovae

arXiv:2605.13983 · astro-ph.IM, astro-ph.HE, hep-ph · Submitted 2026-05-13 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: We're looking at the paper: "Rapid and robust simulation-based inference for kilonovae".

Jocelyn: Here's the gist of the paper. Rapid and robust simulation-based inference for kilonovae develops a simulation-based inference (SBI) framework using a Gaussian process emulator trained on simulations to provide rapid and robust parameter estimation for kilonova data The Problem with Traditional Methods Traditional Bayesian parameter estimation using Markov chain Monte Carlo (MCMC) is time-consuming and relies on explicit likelihood approximations that can break down when modeling uncertainties are significant The standard Gaussian likelihood approximation fails to capture the non-Gaussian, correlated structure of emulator uncertainty The analytical likelihood in standard MCMC approaches leads to biased and overconfident posteriors due to likelihood misspecification The SBI Framework Development The authors develop a simulation-based inference (SBI) framework for kilonova parameter estimation using density-estimation likelihood-free inference This framework uses a Gaussian process emulator trained on ∼ 1300 possis simulations The SBI method accurately recovers injected parameters, while the MCMC suffers from systematic bias caused by likelihood misspecification The GP Emulator Construction A GP emulator is developed to enable SBI to be applied to a broader range of kilonova models The emulator is trained on the model grid described in Section 4.1 The hyperparameters are optimized using GPyTorch (J.

Vera: Gardner et al.

Jocelyn: 2018) The custom batched GP in PyTorch is implemented to enable rapid generation of multi-band light curves Performance Comparison and Robustness The SBI framework provides a rapid alternative to MCMC that is robust to likelihood misspecification The ANPE framework generally produces broader posterior distributions than the likelihood-based analysis, as was the case with the simulation recovery The ANPE framework learns the predictive distribution from forward simulations and does not depend explicitly on residuals or uncertainties, making it more robust to parameterdependent model misspecification Astrophysical Interpretation of AT2017gfo The ANPE infers shape parameters qwind = 83.994+5.18−10.42 and qdyn = 23.718+27.134−13.5, suggesting that the ejecta are not toroidal The wind ejecta from the accretion disk and the dynamical ejecta generated during merger have distinct timescales and compositions The ANPE infers shape parameters qwind = 84 and qdyn = 24, suggesting that the ejecta are not toroidal An SBI framework, once trained, can be applied to new observations without retraining, enabling rapid characterization of kilonova properties across diverse observational campaigns This computational efficiency, combined with robustness to emulator uncertainty and likelihood misspecification, positions SBI as a promising method for maximizing the scientific value of future kilonova observations Author Contributions SMB: Conceptualization, Data Curation, Formal Analysis, Investigation, Methodology, Project Administration, Software, Visualization MB: Conceptualization, Investigation, Methodology, Software, Visualization HVP: Conceptualization, Formal Analysis, Funding Acquisition, Investigation, Methodology NS: Conceptualization, Data Curation, Formal Analysis, Investigation DM: Formal Analysis ST: Methodology, Software GJ: Software SR: Conceptualization, Funding Acquisition, Resources SN: Conceptualization, Writing (original draft) Acknowledgements We thank Justin Alsing, James Alvey, and Oleg Korobkin for valuable conversations SMB is supported by the research project grant “Fundamental physics from populations of compact object mergers” funded by VR under Dnr 2021-04195 MB acknowledges the Department of Physics and Earth Science of the University of Ferrara for the financial support through the FIRD 2025 grant HVP and ST have been supported by funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programmes NS is supported by the Kavli Foundation SR has been supported by the European Research Council (ERC) Advanced Grant INSPIRATION under the European Union’s Horizon 2020 research and innovation programme SN acknowledges support from the Netherlands Organization for Scientific Research (NWO) This research utilized the Sunrise HPC facility supported by the Technical Division at the Department of Physics, Stockholm University Software affine25; corner (D.

Vera: Foreman-Mackey 2016); flowfusion26 (J.

Paper summary: Vera: So we've seen that this paper, "Rapid and robust simulation-based inference for kilonovae," focuses on developing a simulation-based inference framework to estimate kilonova parameters quickly and reliably.

Jocelyn: The core thesis is that traditional MCMC methods are too slow because they rely on explicit likelihood approximations, which can fail when modeling uncertainties are significant >

Subrahmanyan: They propose using density estimation likelihood-free inference, which involves training a neural density estimator, like a normalizing flow or diffusion model, using forward simulations of data parameter pairs >

Vera: And to make this practical for kilonovae, they build a Gaussian process emulator trained on about one thousand three hundred POSSIS simulations to speed things up and handle the complexity >

Jocelyn: The key claim is that once you have this trained framework, like the amortized neural posterior estimator or ANPE, you can apply it to new data rapidly without having to evaluate an explicit likelihood function again >

Subrahmanyan: They show that this SBI approach accurately recovers injected parameters and is robust against likelihood misspecification because it learns the structure of uncertainty directly from the forward simulations >

Vera: So, in short, they’re trading slow MCMC for a fast, robust method that handles those messy uncertainties better than standard likelihood methods >

Jocelyn: It really shows how simulation-based inference is becoming an important tool in this field when we're dealing with complex astrophysical transients like these kilonovae >

Conclusion: Vera: Looking at the whole paper, "Rapid and robust simulation-based inference for kilonovae," it boils down to this: SBI provides a way to get computationally efficient posteriors for kilonovae that are more robust against emulator uncertainty and likelihood misspecification than traditional likelihood-based methods >

Jocelyn: And the authors show that once you have this framework trained, you can apply it to new observations without needing to retrain the whole thing, which is a huge deal for future observation campaigns >

Subrahmanyan: The implications are that we can start rapidly characterizing kilonova properties across diverse observational campaigns with this computational efficiency, combined with robustness to those model misspecifications >

Vera: So when we look at future facilities like LSST or next-generation gravitational-wave detectors, where the diversity of events is going to be huge, this SBI framework positions itself as a promising method for maximizing the scientific value of those observations >

Oskar Klein Centre for Cosmoparticle Physics, Stockholm University · Gravitation Astroparticle Physics Amsterdam (GRAPPA), University of Amsterdam Department of Physics and Earth Science, University of Ferrara Department of Physics and Earth Science, INFN Sezione di Ferrara, INAF Osservatorio Astronomico d’Abruzzo, Institute of Astronomy and Kavli Institute for Cosmology, University of Cambridge Cavendish Laboratory Department of Physics, University Cambridge Astrophysics Group Imperial College London Department of Mathematics Imperial College London Research Computing Services University Cambridge Hamburg Observatory Department of Physics University Hamburg Oskar Klein Centre for Cosmoparticle Physics Stockholm University Deutsches Elektronen-Synchrotron DESY Deutsches Zentrum f¨ur Astrophysik DZA Institut f¨ur Physik und Astronomie Universit¨at Potsdam

astro-ph.IM, astro-ph.HE, hep-ph

Submitted: 2026-05-13

Updated: 2026-11-06

Comments: 27 pages, 13 figures, 6 tables, accepted by APJ

Code: https://github.com/nikhil-sarin/redback

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 73/100

The gist: Rapid and robust simulation-based inference for kilonovae develops a simulation-based inference (SBI) framework using a Gaussian process emulator trained on simulations to provide rapid and robust…

Key concepts

Simulation-Based Inference (SBI)
A method that uses a machine learning model, specifically a Gaussian process emulator trained on simulations, to perform parameter estimation directly from data. Instead of relying on explicit likelihood functions that might be wrong, SBI learns the underlying distribution of parameters directly from the simulated data.
Gaussian Process Emulator (GP Emulator)
A surrogate model trained on a large set of kilonova simulations ($\sim 1300$ POSSIS simulations) to quickly approximate the complex likelihood function. This emulator allows SBI to operate much faster than traditional MCMC by replacing slow likelihood calculations with fast emulator evaluations.
Likelihood Misspecification
A problem in traditional Bayesian analysis where the assumed mathematical form of the probability distribution (the likelihood) does not match the true underlying physical reality. This mismatch causes standard methods to produce biased or overconfident results, which SBI is designed to overcome.
Kilonova Parameter Estimation
The process of determining physical properties of a kilonova event, such as the total ejecta mass and its geometry (e.g., toroidal vs. peanut shape). The paper uses SBI to rapidly and accurately infer these crucial astrophysical characteristics from observational data.

Terminology

Summary

Rapid and robust simulation-based inference for kilonovae develops a simulation-based inference (SBI) framework using a Gaussian process emulator trained on simulations to provide rapid and robust parameter estimation for kilonova data

The Problem with Traditional Methods

Traditional Bayesian parameter estimation using Markov chain Monte Carlo (MCMC) is time-consuming and relies on explicit likelihood approximations that can break down when modeling uncertainties are significant The standard Gaussian likelihood approximation fails to capture the non-Gaussian, correlated structure of emulator uncertainty The analytical likelihood in standard MCMC approaches leads to biased and overconfident posteriors due to likelihood misspecification

The SBI Framework Development

The authors develop a simulation-based inference (SBI) framework for kilonova parameter estimation using density-estimation likelihood-free inference This framework uses a Gaussian process emulator trained on ∼ 1300 possis simulations The SBI method accurately recovers injected parameters, while the MCMC suffers from systematic bias caused by likelihood misspecification

The GP Emulator Construction

A GP emulator is developed to enable SBI to be applied to a broader range of kilonova models The emulator is trained on the model grid described in Section 4.1 The hyperparameters are optimized using GPyTorch (J. p. 5 Gardner et al. p. 1 2018) The custom batched GP in PyTorch is implemented to enable rapid generation of multi-band light curves

Performance Comparison and Robustness

The SBI framework provides a rapid alternative to MCMC that is robust to likelihood misspecification The ANPE framework generally produces broader posterior distributions than the likelihood-based analysis, as was the case with the simulation recovery The ANPE framework learns the predictive distribution from forward simulations and does not depend explicitly on residuals or uncertainties, making it more robust to parameterdependent model misspecification

Astrophysical Interpretation of AT2017gfo

The ANPE infers shape parameters qwind = 83.994+5.18−10.42 and qdyn = 23.718+27.134−13.5, suggesting that the ejecta are not toroidal The wind ejecta from the accretion disk and the dynamical ejecta generated during merger have distinct timescales and compositions The ANPE infers shape parameters qwind = 84 and qdyn = 24, suggesting that the ejecta are not toroidal

Conclusion

An SBI framework, once trained, can be applied to new observations without retraining, enabling rapid characterization of kilonova properties across diverse observational campaigns This computational efficiency, combined with robustness to emulator uncertainty and likelihood misspecification, positions SBI as a promising method for maximizing the scientific value of future kilonova observations

Author Contributions

SMB: Conceptualization, Data Curation, Formal Analysis, Investigation, Methodology, Project Administration, Software, Visualization MB: Conceptualization, Investigation, Methodology, Software, Visualization HVP: Conceptualization, Formal Analysis, Funding Acquisition, Investigation, Methodology NS: Conceptualization, Data Curation, Formal Analysis, Investigation DM: Formal Analysis ST: Methodology, Software GJ: Software SR: Conceptualization, Funding Acquisition, Resources SN: Conceptualization, Writing (original draft)

Acknowledgements

We thank Justin Alsing, James Alvey, and Oleg Korobkin for valuable conversations SMB is supported by the research project grant “Fundamental physics from populations of compact object mergers” funded by VR under Dnr 2021-04195 MB acknowledges the Department of Physics and Earth Science of the University of Ferrara for the financial support through the FIRD 2025 grant HVP and ST have been supported by funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programmes NS is supported by the Kavli Foundation SR has been supported by the European Research Council (ERC) Advanced Grant INSPIRATION under the European Union’s Horizon 2020 research and innovation programme SN acknowledges support from the Netherlands Organization for Scientific Research (NWO) This research utilized the Sunrise HPC facility supported by the Technical Division at the Department of Physics, Stockholm University

Software

affine25; corner (D. Foreman-Mackey 2016); flowfusion26 (J. p. 2 Alsing et al. p. 2 2024; S. Thorp et al. p. 0 2024a, 2025; B. p. 1 Leistedt et al. 2026); GPyTorch 25 https://github. p. 2com/justinalsing/affine

Data Availability

The AT2017gfo data is publicly available through the Open Access Catalog API (J. Guillochon et al. p. 2 2017)

References

Abac, A., Abramo, R., Albanesi, S., et al. 2026, JCAP, 2026(03), 081 <ref:2605. p. 013983#pg2> Abbott, B. p. 0 P., Abbott, R. p. 0 p. 0 p. 0 p. 0, Abbott, T. p. 0 p. 0 p. 0 p. 0 D., et al. 2017a, PhRvL, 119 <ref:2605. p. 013983#pg5> Abbott, B. p. 0 p. 0 p. 0 P., Abbott, R., Abbott, T. D., et al. 2017b, ApJL <ref:2605. p. 013983#pg5> Abbott, B. P., Abbott, R., Abbott, T. D., et al. 2017c, Nature <ref:2605. p. 013983#pg5> Abbott, B. P., Abbott, R., Abbott, T. D., et al. 2017d, ApJL <ref:2605. p. 013983#pg5> Ajello, M. p. 0, Allafort, A., Axelsson, M., et al. 2018, ApJ <ref:2605. p. 013983#pg3> Almualla, M., Ning, Y., Bulla, M. p. 0 p. 0, et al. 2022 <ref:2605. p. 0 p. 0 p. 013983#pg2> Alsing, J. p. 2, Charnock, T., Feeney, S., & Wandelt, B. 2019, MNRAS <ref:2605. p. 0 p. 013983#pg5> Anand, S. p. 1, Pang, P. T. H., Bulla, M., et al. 2023 <ref:2605. p. 0 p. 0 p. 013983#pg2> Andreoni, I. p. 1, Ackley, K., Cooke, J., et al. 2017, PASA <ref:2605. p. 013983#pg3> Andreoni, I. p. 1, Coughlin, M. p. 1 p. 1 W., Almualla, M., et al. 2021 <ref:2605. p. 013983#pg5> Annala, E. p. 0, Gorda, T., Kurkela, A., & Vuorinen, A. 2018, PhRvL <ref:2605. p. 013983#pg3> Ansel, J., Yang, E., He, H., et al. 2024 <ref:2605. p. 0 p. 0 p. 0 p. 013983#pg5> Arcavi, I. p. 2, Hosseinzadeh, G., Howell, D. A., et al. 2017, Nature <ref:2605. p. 013983#pg4> Arruda, J. p. 2, Bracher, N., K¨othe, U., Hasenauer, J., & Radev, S. T. 2025 <ref:2605. p. 0 p. 0 p. 0 p. 013983#pg5> Ashton, G. p. 1, Bernstein, N., Buchner, J., et al. 2022 <ref:2605.13983#pg5> Auer, P. p. 5 2002, JMLR <ref:2605. p. 013983#pg9> Auer, P. p. 5, Cesa-Bianchi, N., & Fischer, P. 2002, Machine Learning <ref:2605. p. 013983#pg9> Banerjee, S. p. 4, Tanaka, M. p. 1 p. 1, Kato, D., et al. 2022 <ref:2605.13983#pg6> Banerjee, S. p. 4, Tanaka, M., Kawaguchi, K. p. 4, Kato, D., & Gaigalas, G. 2020 <ref:2605. p. 0 p. 0 p. 0 p. 013983#pg6> Barbary, K. p. 4, Bailey, S., Barentsen, G., et al. 2025 <ref:2605.13983#pg12> Barnes, J. p. 3, Kasen, D. p. 1 p. 1, Wu, M.-R., & Mart´ınez-Pinedo, G. 2016 <ref:2605. p. 013983#pg6> Bauswein, A. p. 0, Goriely, S., & Janka, H.-T. 2013 <ref:2605. p. 013983#pg4> Bauswein, A. p. 0, Just, O., Janka, H.-T., & Stergioulas, N. 2017 <ref:2605. p. 0 p. 0 p. 0 p. 0 p. 0 p. 013983#pg4> Bellm, E. p. 2 C., Kulkarni, S. p. 1 R., Graham, M. J., et al. 2018 <ref:2605. p. 0 p. 0 p. 0 p. 013983#pg3> Biewald, L. 2020, Experiment Tracking with Weights and Biases <ref:2605. p. 013983#pg12> Bird, S. p. 4, Rogers, K. p. 4 K., Peiris, H. p. 0 V., et al. 2019, JCAP <ref:2605. p. 013983#pg5> Bovard, L., Martin, D., Guercilena, F., et al. 2017 <ref:2605.13983#pg4> Brethauer, D. p. 4, Kasen, D., Margutti, R., & Chornock, R. 2024 <ref:2605.13983#pg6> Bulla, M. p. 0 2019, MNRAS <ref:2605.13983#pg5> Bulla, M. p. 0 p. 0 2023, MNRAS <ref:2605. p. 013983#pg5> Bulla, M., Sim, S. A., & Kromer, M. 2015 <ref:2605. p. 013983#pg5> Cantiello, M. p. 2, Jensen, J. B., Blakeslee, J. P., et al. 2018 <ref:2605.13983#pg3> Capano, C. p. 0 D., Tews, I., Brown, S. p. 0 M., et al. 2019 <ref:2605. p. 013983#pg5> Chen, R. T. Q., Rubanova, Y., Bettencourt, J., & Duvenaud, D. K. 2018 <ref:2605.13983#pg6> Collins, C. p. 1 E., Shingles, L. J., Bauswein, A. p. 0, et al. 2024 <ref:2605.13983#pg6> Colombo, A. p. 1, Salafia, O. S., Ghirlanda, G., et al. 2025 <ref:2605.13983#pg12> Combi, L., & Siegel, D. M. 2023 <ref:2605.13983#pg12> Coughlin, M. p. 1 W., Dietrich, T. p. 1 p. 1, Doctor, Z., et al. 2018 <ref:2605.13983#pg5> Coughlin, M. p. 1 W., Dietrich, T., Heinzel, J., et al. 2020 <ref:2605.13983#pg5> Coulter, D. p. 0 A., Foley, R. J., Kilpatrick, C. D., et al. 2017 <ref:2605.13983#pg3> Cowperthwaite, P. p. 1 S., Berger, E., Villar, V. p. 0 A., et al. 2017 <ref:2605.13983#pg3> Cranmer, K. p. 2, Brehmer, J., & Louppe, G. 2020 <ref:2605.13983#pg15> da Costa-Luis, C., Larroque, S. K., Altendorf, K., et al. 2026 <ref:2605. p. 013983#pg14> Dani, V. p. 5, Hayes, T. P., & Kakade, S. M. 2008 <ref:2605. p. 013983#pg14> Darc, P. p. 1, Bom, C. R., Fraga, B. M. O., & Kilpatrick, C. D. 2024 <ref:2605.13983#pg14> Desai, M. p. 1 M., Chatterjee, D., Jhawar, S. p. 1, et al. 2025 <ref:2605.13983#pg14> D´ıaz, M. p. 2 C., Macri, L. M., Garcia Lambas, D., et al. 2017 <ref:2605.13983#pg3> Dietrich, T. p. 1, Coughlin, M. W., Pang, P. T. H., et al. 2020 <ref:2605.13983#pg5> Cormerkerkerer J.-P and the likelihood-based analysis leads to biased and overconfident posteriors due to likelihood misspecification; this effect is amplified near prior boundaries <ref:2605. p. 213983#pg20>

Figure 7

Figure 7 shows posterior bias assessment for the kilonova total mass <ref:2605.13983#pg14> The right-hand panel shows that the average biases are −0.065σ and −0.05σ, respectively, indicating that the MAP predictions are well within the 1σ posterior bounds <ref:2605.13983#pg14>

Figure 8

Figure 8 shows posterior distributions from a simulation recovery using an MCMC sampler (purple) and the ANPE (grey) <ref:2605.13983#pg16> The right-hand panel shows that the average biases are −0.065σ and −0.05σ, respectively, indicating that the MAP predictions are well within the 1σ posterior bounds <ref:2605.13983#pg14>

Figure 9

Figure 9 shows posterior predictive light curves for a simulated light curve, comparing inference using MCMC (left) and ANPE (right) <ref:2605.13983#pg19> The bottom two panels show the distribution of the flux–to-uncertainty ratio for the reddest bands (JHK) and for the remaining optical bands (ugrizy) <ref:2605. p. 913983#pg19>

Table 3

Parameter bias recovery summary for model parameters excluding masses, which are shown in Fig. 7 <ref:2605.13983#pg15> For all parameters except dL, the absolute mean normalized bias is less than 1, indicating no significant systematic bias <ref:2605. p. 813983#pg15>

Table 4

Parameter estimation summary for AT2017gfo <ref:2605. p. 213983#pg16> In the likelihood-based analysis, several parameters are driven toward the boundaries of the prior space, including dL, θv, vwind ej and qwind <ref:2605. p. 0 p. 013983#pg16> The pile up of samples at the prior-boundary seen in the dL posterior in the simulation recovery (Section 7. p. 0 p. 01) now occurs for multiple parameters <ref:2605. p. 0 p. 013983#pg16>

Discussion

The ANPE framework generally produces broader posterior distributions than the likelihood-based analysis, as was the case with the simulation recovery <ref:2605.13983#pg20> This reflects both the incorporation of the full GP predictive covariance and the over-estimate of uncertainty shown in Section 6. p. 63 <ref:2605.13983#pg20> The ANPE framework, on a single NVIDIA HGX A100 GPU card, can generate ∼ 2 × 104 posterior samples in approximately 4 seconds for a single observation <ref:2605. p. 0 p. 013983#pg20> All analyses were run using a single NVIDIA HGX A100 GPU card <ref:2605.13983#pg20> The ANPE framework, on a single NVIDIA HGX A100 GPU card, can generate ∼ 2 × 104 posterior samples in approximately 4 seconds for a single observation <ref:2605.13983#pg20>

The results in this paper demonstrate that SBI can provide computationally efficient posteriors for kilonovae that are more robust to emulator uncertainty and likelihood misspecification than traditional likelihoodbased methods <ref:2605.13983#pg22> The posterior predictive light curves show that the ANPE reasonably reproduces observed AT2017gfo light curves under the assumed model <ref:2605. p. 913983#pg22> As kilonova discovery rates accelerate with new facilities such as JWST, Roman, LSST, and next-generation gravitational-wave detectors, the diversity of observed events will span broad regions of parameter space with varying merger geometries, electron fractions, and mass ratios <ref:2605.13983#pg22> An SBI framework, once trained, can be applied to new observations without retraining <ref:2605. p. 213983#pg22> This computational efficiency, combined with robustness to emulator uncertainty and likelihood misspecification, positions SBI as a promising method for maximizing the scientific value of future kilonova observations <ref:2605. p. 013983#pg22>

Table 5

Table 5. Notation and symbols used throughout the paper The parameters are organized by category kilonova models, likelihood-based Bayesian inference, Gaussian process emulation, and simulation based inference with flow matching

Figure 10

Figure 10 shows posterior distributions from AT2017gfo using an MCMC sampler (purple) and the ANPE (grey) <ref:2605.13983#pg26> The likelihood-based ones, which tend to overestimate the brightness (consistent with Fig. p. 8 6), the ANPE predictions also provide better late-time matches in the y, J, and H bands <ref:2605. p. 113983#pg26>

Figure 11

Figure 11 shows posterior predictive light curves for both analyses <ref:2605.13983#pg27> The wind (blue) and dynamical (red) ejecta components are shown separately, with the total luminosity shown in black <ref:2605. p. 6 p. 613983#pg27> The wind component takes over between 1–5 days depending on wavelength <ref:2605.13983#pg27>

Figure 12

Figure 12 shows predicted bolometric light curves for AT2017gfo based on the ANPE (left) and MCMC (right) posteriors <ref:2605.13983#pg27> The wind (blue) and dynamical (red) ejecta components are shown separately, with the total luminosity shown in black <ref:2605.13983#pg27>

Table 4. Parameter estimation summary for AT2017gfo <ref:2605.13983#pg16> In the likelihood-based analysis, several parameters are driven toward the boundaries of the prior space, including dL, θv, vwind ej and qwind <ref:2605.13983#pg16> The pile up of samples at the prior-boundary seen in the dL posterior in the simulation recovery (Section 7.1) now occurs for multiple parameters <ref:2605.13983#pg16>

Figure 8 shows posterior distributions from a simulation recovery using an MCMC sampler (purple) and the ANPE (grey) <ref:2605.13983#pg16> The true parameter values are shown in red <ref:2605. p. 313983#pg16>

Figure 6

Figure 6 shows the resulting sampling distribution in representative bands at observation time t = 2.5 d <ref:2605. p. 013983#pg11> Some of the empirical distributions have noticeable asymmetry relative to a Gaussian distribution <ref:2605.13983#pg11> This is what the ANPE is trained on <ref:2605. p. 013983#pg11>

Figure 5

Figure 5 shows this distribution around tobs = 2.5 d in the u, g, and H bands <ref:2605. p. 013983#pg12> Some of the empirical distributions have noticeable asymmetry relative to a Gaussian distribution <ref:2605.13983#pg12> The Gaussian model over-predicts the probability of faint flux values and under-predicts bright fluxes when compared to the empirical sampling distribution <ref:2605. p. 9

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved inference speed for kilonova parameters by enabling rapid generation of posterior samples in seconds, compared to hours for likelihood-based inference, because the neural posterior estimator can generate 2 × 104 posterior samples in approximately 4 seconds for a single observation whereas the likelihood-based analysis requires roughly 9 hours (for 1000 walkers and 1400 iterations).

  2. Robust parameter estimation by using density-estimation likelihood-free inference that avoids biased or overconfident posterior predictions caused by likelihood misspecification, as it learns the conditional density P(Phid) directly from forward simulations.

  3. Capture of complex emulator uncertainty in parameter space by using a Gaussian process emulator trained on 1200 single-component possis simulations and augmenting it with Bayesian Emulator Optimization (BEO) to improve model accuracy, which is shown to reduce the median ratio of emulator uncertainty to observational uncertainty by an average of 24%.

  4. Enhanced parameter recovery by using simulation-based inference that sidesteps likelihood misspecification by learning the posterior distribution directly from forward simulations, making it more robust to large and complex emulator uncertainties that are not captured by a diagonal Gaussian likelihood.

  5. Improved geometric constraint on ejecta structure because the SBI framework infers shape parameters q wind = 83.94+5.18−10.42 and q dyn = 23.5+25.5−13.5, suggesting that the toroidal (q = 0) and peanut (q = 1.5) geometries lie outside the 99th percentile credible intervals for both components.

  6. Better identification of systematic biases in distance and dynamical ejecta mass by comparing the ANPE to likelihood-based analysis, showing that the inferred luminosity distance is railing against the upper boundary of the prior range and away from the true value in likelihood-based analyses.

Sources

Related papers