Source-linked AI summary
Real-time gravitational-wave science with neural posterior estimation
Maximilian Dax, Stephen R. Green, Jonathan Gair, Jakob H. Macke, Alessandra Buonanno, Bernhard Schölkopf
TL;DR
Rapid gravitational-wave inference must match standard posterior estimates while avoiding their long runtimes. DINGO uses neural posterior estimation trained on simulated signals and detector-noise PSDs, and its analyses show extremely close agreement with standard samplers, including for real events, while addressing PSD-dependent effects and known scope limitations.
Problem
The paper addresses the need for accurate gravitational-wave parameter inference across the full 15D binary-black-hole space while standard sampling comparisons remain computationally demanding.
Method
DINGO trains neural posterior estimators on simulated frequency-domain signals with detector-noise PSDs and compares their posteriors with LALInference and Bilby standard samplers.
Results
90% credible intervals are in extremely close agreement with standard samplers, while PSD mismatches produce larger divergences than DINGO–LALInference comparisons.
Takeaways & Limitations
Conditioning the neural network on the event-specific PSD is required at the accuracy level achieved in this study.
Takeaways & Limitations
Agreement is theoretically guaranteed only for data consistent with the training distribution, whereas real detector noise is neither perfectly stationary nor Gaussian.
Abstract
from arXiv · showhide
We demonstrate unprecedented accuracy for rapid gravitational-wave parameter estimation with deep learning. Using neural networks as surrogates for Bayesian posterior distributions, we analyze eight gravitational-wave events from the first LIGO-Virgo Gravitational-Wave Transient Catalog and find very close quantitative agreement with standard inference codes, but with inference times reduced from O(day) to a minute per event. Our networks are trained using simulated data, including an estimate of the detector-noise characteristics near the event. This encodes the signal and noise models within millions of neural-network parameters, and enables inference for any observed data consistent with the training distribution, accounting for noise nonstationarity from event to event. Our algorithm -- called "DINGO" -- sets a new standard in fast-and-accurate inference of physical parameters of detected gravitational-wave events, which should enable real-time data analysis without sacrificing accuracy.
TRAINING DATA
DINGO is trained on simulated 15-dimensional binary-black-hole data paired with detector-noise PSDs, using distance-range partitions and observing-run noise estimates. The training construction combines saved intrinsic waveforms with sampled extrinsic parameters and noise realizations.
- TRAINING DATA: DINGO performs inference over the full 15D parameter space for quasicircular binary black holes.The parameters include masses, coalescence time, phase, sky position, distance, inclination, spin, and polarization quantities.
- TRAINING DATA: Accurate training across luminosity distance requires partitioning the prior into distance ranges.The authors report difficulty training one network over the entire relevant distance range.
- TRAINING DATA: Training data pair labeled strain sets with associated noise power spectral densities sampled from empirical observing-run distributions.PSD estimates are constructed from event-free detector stretches and used similarly before real events.
- TRAINING DATA: Each simulated strain set follows d(i) = h(θ(i)) + n(i), using frequency-domain IMRPhenomPv2 waveforms with additive stationary Gaussian noise.The data span 20–1024 Hz at 0.125 Hz resolution, corresponding to 8 seconds.
- TRAINING DATA: The waveform construction uses 5 × 10^6 intrinsic-parameter sets, while extrinsic parameters and noise realizations are sampled during training.This makes the effective training-set size infinite in those dimensions.
NEURAL NETWORK
The DINGO inference network has two main components: an embedding network that compresses detector data into features and a normalizing flow that produces the Bayesian posterior.
- NEURAL NETWORK: The embedding network compresses observed detector data into a sufficiently small number of features.This compression supplies the context used by the posterior estimator.
- NEURAL NETWORK: The normalizing flow transforms the extracted features into a Bayesian posterior distribution.It is the density-estimation component that produces posterior samples for parameter inference.
- NEURAL NETWORK: Together, the embedding network and normalizing flow form DINGO’s conditional density-estimation neural network.The two components divide the task between data representation and posterior generation.
Embedding network
The embedding network processes whitened frequency-domain strain and inverse ASD inputs for each detector, first reducing their dimensionality and then refining the representation with residual blocks.
- Embedding network: 24,096 inputs per detector comprise real and imaginary whitened strain components plus the inverse amplitude spectral density.The inverse ASD is used for more numerically stable behavior around spectral lines.
- Embedding network: The first embedding layer linearly reduces the detector data to 400 components per detector.This layer is initialized using singular-value decomposition to provide an inductive bias.
- Embedding network: The strain and inverse-ASD inputs are projected onto 200 complex singular vectors, yielding 800 or 1200 real features for two or three detectors.The inverse-ASD projection is added to the projected strain, and no nonlinear activation follows this layer.
- Embedding network: A fully connected residual network follows the SVD layer with 48 hidden layers across blocks of 1024, 512, 256, and 128 dimensions.The residual blocks use batch normalization and ELU activations before each linear mapping.
Normalizing flow
DINGO’s normalizing flow uses coupling transforms conditioned on embedded detector features and perturbed coalescence times, alongside a neural initial estimate for detector arrival times.
- Normalizing flow: The flow uses rational-quadratic spline coupling transforms that update half of a sample’s components using context and the remaining components.The transform structure is similar to the earlier approach but changes the residual-block count.
- Normalizing flow: 30 coupling transforms increase the flow depth from 15 in the earlier study, producing 300 hidden layers.The combined networks contain 1.31 × 10^8 parameters for two detectors and 1.42 × 10^8 for three detectors.
- Normalizing flow: Flow context consists of 128 embedding features and two or three perturbed detector coalescence times τI.These perturbed times provide detector-specific timing context to the posterior estimator.
- Normalizing flow: A separate neural network supplies initial estimates of detector coalescence times for the iterative GNPE algorithm.This initial-estimate network has the same form as the main network but does not use τI as context.
Training
Training uses a two-stage procedure that begins with fixed average noise characteristics and then fine-tunes on empirically varying detector noise. The loss behavior reflects the broader training distribution introduced during fine-tuning.
- Training consists of pretraining with a frozen SVD layer and average-PSD noise, followed by fine-tuning with an unfrozen SVD layer and empirically sampled PSDs.
- The loss jumps when fine-tuning begins because varying noise PSDs broaden the training-data distribution.
- 300 pretraining epochs and 150 fine-tuning epochs are used with batch size 4096, while three-detector training reduces the batch size to 2048 because of memory limitations.
EFFECT OF PSD
DINGO conditions inference on detector-noise PSDs, and using an incorrect PSD produces substantially larger posterior discrepancies, especially for distance, sky position, and coalescence time.
- 0.005 nat is the mean JSD and 0.030 nat the maximum JSD when GW150914 is evaluated with GW151012’s PSD instead of the correct PSD.
- 0.001 nat is the largest mean JSD and 0.006 nat the largest maximum JSD between DINGO and LALInference across the reported events and parameters.
- Conditioning the neural network on the PSD is required at the accuracy level achieved in this study.
- The incorrect PSD most affects tc, dL, α, and δ, which have a mean JSD of 0.020 nat.The passage attributes these biases to incorrect inferred signal amplitudes, affecting distance directly and sky position and coalescence time through relative amplitudes.
COMPARISONS AGAINST STANDARD SAMPLERS
DINGO is compared with standard samplers using matched data conditioning and achieves very close agreement across gravitational-wave posterior estimates. The comparison also identifies convergence and out-of-distribution noise as sources of discrepancy.
- COMPARISONS AGAINST STANDARD SAMPLERS: Standard samplers marginalize selected parameters to improve sampling performance, reconstructing marginalized parameters during postprocessing.
- COMPARISONS AGAINST STANDARD SAMPLERS: DINGO and LALInference MCMC produce 90% credible intervals that are all in extremely close agreement across the reported parameters.
- COMPARISONS AGAINST STANDARD SAMPLERS: DINGO agrees most closely with LALInference MCMC, which is therefore used for the main-text comparisons and posterior plots.
- COMPARISONS AGAINST STANDARD SAMPLERS: 0.0015 nat and 0.0010 nat are the mean JSDs between DINGO and Bilby and between DINGO and LALInference nested sampling, respectively, across events and parameters.
- COMPARISONS AGAINST STANDARD SAMPLERS: DINGO deviations can arise from incomplete network convergence, imperfect standard-sampler samples, and data whose noise is outside the training distribution.For nonstationary or non-Gaussian noise, the paper states that no theoretical guarantee ensures DINGO extrapolates like standard samplers.
- COMPARISONS AGAINST STANDARD SAMPLERS: The posterior plots include marginalized one- and two-dimensional distributions for GW150914 and separate figures for the listed GWTC-1 events, including the only three-detector event, GW170814.