[Paper Review] Statistically-informed deep learning for gravitational wave parameter estimation

Hongyu Shen, E. A. Huerta+3 more|arXiv (Cornell University)|Mar 5, 2019

Pulsars and Gravitational Waves ResearchPhysics and Astronomy84 references26 citations

TL;DR

This paper introduces a statistically-informed deep learning framework that estimates gravitational wave parameters—binary black hole masses (m1, m2), final spin (af), and ringdown mode frequencies and damping times (ωR, ωI)—directly from time-series LIGO strain data. By combining a modified WaveNet architecture with contrastive learning and normalizing flows, the model produces posterior distributions consistent with Bayesian inference, achieving sub-millisecond inference on a single V100 GPU while matching PyCBC Inference results on five real events.

ABSTRACT

We introduce deep learning models to estimate the masses of the binary components of black hole mergers, $(m_1,m_2)$, and three astrophysical properties of the post-merger compact remnant, namely, the final spin, $a_f$, and the frequency and damping time of the ringdown oscillations of the fundamental $\ell=m=2$ bar mode, $(\omega_R, \omega_I)$. Our neural networks combine a modified $ exttt{WaveNet}$ architecture with contrastive learning and normalizing flow. We validate these models against a Gaussian conjugate prior family whose posterior distribution is described by a closed analytical expression. Upon confirming that our models produce statistically consistent results, we used them to estimate the astrophysical parameters $(m_1,m_2, a_f, \omega_R, \omega_I)$ of five binary black holes: $ exttt{GW150914}, exttt{GW170104}, exttt{GW170814}, exttt{GW190521}$ and $ exttt{GW190630}$. We use $ exttt{PyCBC Inference}$ to directly compare traditional Bayesian methodologies for parameter estimation with our deep-learning-based posterior distributions. Our results show that our neural network models predict posterior distributions that encode physical correlations, and that our data-driven median results and 90$\%$ confidence intervals are similar to those produced with gravitational wave Bayesian analyses. This methodology requires a single V100 $ exttt{NVIDIA}$ GPU to produce median values and posterior distributions within two milliseconds for each event. This neural network, and a tutorial for its use, are available at the $ exttt{Data and Learning Hub for Science}$.

Motivation & Objective

• To develop a computationally efficient deep learning model for estimating key astrophysical parameters of binary black hole mergers from gravitational wave strain data.
• To ensure statistical consistency by training against a Gaussian conjugate prior with analytically tractable posteriors.
• To estimate not only masses but also post-merger remnant properties: final spin (af), and quasinormal mode frequencies and damping times (ωR, ωI).
• To validate the model’s predictive accuracy and physical consistency against established Bayesian inference via PyCBC Inference.
• To enable real-time, high-throughput parameter estimation for upcoming large-scale gravitational wave surveys.

Proposed method

• A modified WaveNet architecture processes time-series gravitational wave strain data to extract features relevant to source parameters.
• Contrastive learning is applied to improve feature representation by encouraging similarity between positive pairs of data and contrast between negative pairs.
• Normalizing flows are used to model complex, multi-modal posterior distributions over the parameters (m1, m2, af, ωR, ωI).
• The model is trained on simulated gravitational wave signals embedded in advanced LIGO noise, using a Gaussian conjugate prior for statistical validation.
• Posterior samples are generated from the learned normalizing flow, enabling uncertainty quantification and credible interval estimation.
• Inference is accelerated using a single V100 GPU, achieving median estimate and full posterior prediction in under 2 milliseconds per event.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on simulated data produce statistically consistent posterior distributions for binary black hole parameters?
• RQ2 How accurately can the model estimate the final spin (af) and ringdown mode parameters (ωR, ωI) directly from time-series strain data?
• RQ3 To what extent do the deep learning-generated posterior distributions encode physical correlations between parameters?
• RQ4 How do the median estimates and 90% credible intervals from the deep learning model compare to those from traditional Bayesian inference (PyCBC Inference)?
• RQ5 Can the model achieve high-accuracy parameter estimation with minimal computational cost, suitable for real-time analysis of large-scale gravitational wave data?

Key findings

• The deep learning model produces posterior distributions that are statistically consistent with the analytical posteriors derived from a Gaussian conjugate prior family.
• For five real gravitational wave events—GW150914, GW170104, GW170814, GW190521, and GW190630—the model's median estimates and 90% credible intervals closely match those from PyCBC Inference.
• The model successfully captures physical correlations between parameters, particularly between the binary masses and the final spin (af).
• The model achieves inference speed of under 2 milliseconds per event on a single V100 GPU, enabling real-time application in large-scale surveys.
• The posterior distributions generated by the model are multi-modal and reflect the true uncertainty structure of the parameter space, as confirmed by comparison with Bayesian methods.
• The framework is publicly released with a tutorial at the Data and Learning Hub for Science, enabling reproducibility and community adoption.