Learning from simulations could make kilonova analysis faster and more reliable

Reading the glow of a merger

A kilonova is a short-lived glow powered by radioactive decay in material expelled during a merger involving neutron stars. Its changing brightness across ultraviolet, optical and near-infrared wavelengths can help researchers estimate how much material escaped, how fast it moved and what it contained.

Those estimates are not direct measurements. They depend on models of how radiation travels through expanding matter. Detailed simulations are expensive, while faster approximations introduce uncertainties that can complicate the interpretation of observations.

In their study, accepted by The Astrophysical Journal and available on arXiv, Brown and colleagues develop a simulation-based inference framework designed to handle these uncertainties without imposing a simple mathematical approximation for their distribution.

Two stages of learning

The researchers first built a Gaussian-process emulator: a computational shortcut that predicts light curves between detailed simulations and estimates its own interpolation uncertainty. Its training set began with 1,200 simulations from the three-dimensional radiation-transport code POSSIS. An optimization procedure added another 124 simulations, concentrating effort on regions relevant to interpreting AT2017gfo, the kilonova associated with GW170817.

The model combines two separately simulated ejecta components. Each has its own mass, composition, velocity and shape. Together with a shared viewing angle and the system’s distance, these give ten parameters to infer.

The second learning stage used one million synthetic datasets generated with the emulator. These included random draws representing emulator uncertainty, plus observational noise matching the AT2017gfo measurements. A neural probability model trained using a technique called flow matching then learned to produce posterior distributions—the ranges of parameter values compatible with data under the adopted model and prior assumptions.

For the real event, the analysis used 620 brightness measurements across 25 photometric bands, spanning approximately 0.45 to 29.4 days after the merger. Independent measurements supplied informative prior constraints on distance and viewing angle.

Why the uncertainty model matters

The comparison analysis used Markov chain Monte Carlo, or MCMC, with the same emulator and priors. Its likelihood—the rule assigning probabilities to observations for a proposed parameter set—approximated the uncertainties as Gaussian and uncorrelated between measurements.

That approximation did not fully reproduce the emulator’s predictive distribution. Transforming the emulator’s outputs into flux can produce asymmetric uncertainties, while correlations also matter. The simulation-trained method learns from draws of that predictive distribution rather than relying on the diagonal Gaussian approximation.

This distinction is important: the paper identifies problems with the chosen likelihood, not an inherent failure of MCMC. MCMC is a sampling method; its conclusions depend on the probability model it samples.

In 50 synthetic-data tests spread across the adopted parameter ranges, the neural method’s mean signed estimation offsets were below one posterior standard deviation in absolute magnitude for every parameter. The likelihood-based comparison exceeded two for distance, both component masses and the slower component’s velocity. Other validation tests suggested that the neural method sometimes produced conservative, overly broad uncertainties. These checks support its performance in the tested settings, rather than establishing universal accuracy.

What the reanalysis suggests

For AT2017gfo, the simulation-based method inferred approximately 0.087 solar masses of ejecta. About 86% belonged to the more massive, relatively lanthanide-poor component. Lanthanides are a family of heavy elements whose presence strongly affects how radiation escapes.

The inferred component velocities were approximately 7.4% and 18.4% of the speed of light. Within the assumed shape model, toroidal and peanut-shaped configurations lay outside the 99% credible intervals for both components. These are model-dependent constraints, not direct images of the debris. The authors also caution against treating the component labels as definitive evidence of their physical origins.

Several likelihood-based estimates accumulated near the limits of their allowed ranges. The neural analysis avoided the same pattern and generally returned broader distributions, with distance and viewing angle remaining largely controlled by their independent priors.

Fast inference, with physical limits

After training, the neural estimator generated about 20,000 posterior samples in roughly four seconds. The comparison MCMC analysis took about nine hours; both ran on a single NVIDIA HGX A100 GPU. That timing excludes the upfront simulation and training work.

Speed does not repair incomplete physics. Both approaches inherit uncertain opacities, omit some possible early-emission processes and combine component light curves without modeling radiation passing between them. Both struggled with the bluest observations. Training also used AT2017gfo’s specific observation times, bands and noise, so performance across different observing arrangements remains to be established.

The arXiv manuscript, version 3, is listed as accepted by The Astrophysical Journal. Its central result is a promising methodological demonstration: learning a more realistic uncertainty distribution can improve inference, while leaving the limitations of the underlying physical model intact.