Rapid Likelihood Free Inference of Compact Binary Coalescences using Accelerated Hardware

TL;DR

AMPLFI delivers real-time, likelihood-free gravitational-wave parameter estimation by combining a GPU-accelerated embedding network with inverse autoregressive flows to learn posterior distributions from BBH simulations in real detector noise. Integrated with Aframe, the framework minimizes latency by keeping data on GPUs, generating waveforms on the fly, and performing fast posterior sampling to produce online data products such as skymaps. The results show AMPLFI achieves low-latency inference with reasonable agreement to traditional Bayesian methods for intrinsic parameters, while extrinsic parameters and sky localization remain more challenging, highlighting the need for targeted improvements and periodic re-training. Overall, this work demonstrates a practical path to online GW alerts and multi-messenger follow-up by leveraging accelerators and likelihood-free inference, with potential extension to broader CBC and non-CBC morphologies.

Abstract

We report a gravitational-wave parameter estimation algorithm, AMPLFI, based on likelihood-free inference using normalizing flows. The focus of AMPLFI is to perform real-time parameter estimation for candidates detected by machine-learning based compact binary coalescence search, Aframe. We present details of our algorithm and optimizations done related to data-loading and pre-processing on accelerated hardware. We train our model using binary black-hole (BBH) simulations on real LIGO-Virgo detector noise. Our model has $\sim 6$ million trainable parameters with training times $\lesssim 24$ hours. Based on online deployment on a mock data stream of LIGO-Virgo data, Aframe + AMPLFI is able to pick up BBH candidates and infer parameters for real-time alerts from data acquisition with a net latency of $\sim 6$s.

Figures (10)

• Figure 1: A comparison of time-domain IMRPhenomD between that implemented in this study, as a part of the ml4gw library. The parameters of the waveform is ${\mathcal{M} = 26\;M_{\odot}, q = 1.0, D_L = 1000\;\mathrm{Mpc}, \chi_{1,2} = 0.0}$. We find that while there are differences in the waveform strain, the residuals are below three orders of magnitude compared to the signal for most of the evolution, except the final few cycles where it is two orders of magnitude lower.
• Figure 2: The figure shows the whitened time-domain background strain from Hanford and Livingston (HL) in two different colors. This is a stretch of data from May 2019 (early O3). A simulated BBH signal is injected; the waveform is overlayed. In the bottom panel, we see the same background strain with a time-shifted signal injected, the parameters of which are otherwise the same as the top panel. The time shift is random up to 1s compared to the top panel, indicated by the shading. We summarize our data views like $d$ and $d^\prime$, and embedding them jointly. Subsequently, for LFI, we only use $d^\prime$-like views.
• Figure 3: Example posterior for a signal with parameters $\{\mathcal{M}=45M_{\odot}, q = 0.7, D_L = 1000~\mathrm{Mpc}\}$ injected in 20 different background instances, sampled using AMPLFI is shown in sky blue . All posteriors are consistent with one another. Posteriors from the same signal injected in 5 different background instances (same background stretch as the AMPLFI injections) and analyzed via nested sampling with Bilby, is overlayed in varying Orange-red colors. We find that parameters like $\mathcal{M}$ and $q$ are consistent in terms of detection uncertainties across different runs. Extrinsic parameters, especially the sky-location, though consistent with the true parameters, shows larger uncertainty with AMPLFI.
• Figure 4: Example posterior for a signal with parameters $\{\mathcal{M}=15M_{\odot}, q = 0.9, D_L = 1000~\mathrm{Mpc}\}$ injected in 20 different background instances, sampled using AMPLFI is shown in sky blue. All posteriors are consistent with one another. Posteriors from the same signal injected in 5 different background instances (same background stretch as the AMPLFI injections) and analyzed via nested sampling with Bilby, is overlayed in varied Orange-red colors. We find that parameters like $\mathcal{M}$ and $q$ are consistent in terms of detection uncertainties across different runs. Extrinsic parameters, especially the sky-location, though consistent with the true parameters, shows larger uncertainty with AMPLFI.
• Figure 5: Left: Percentile-percentile (PP) plot showing recovery accuracy for 500 BBH injections performed in a testing background, different from training background. The different lines track the cumulative fraction of events within a corresponding confidence interval for the parameters mentioned in Eq. \ref{['eq:parameters']}. Right: Sampling times for AMPLFI vs. nested sampling runs done on injections using Bilby, with identical waveform model and prior settings. The nested sampling runs were done with a CPU pool size of 24, and correspond to the runs using in Figure \ref{['fig:low-mass-compare']}. The standard GW likelihood model is used. The AMPLFI sampling times correspond to the 500 injections used for the P-P plot on the left.