3D Binary Neutron Star Merger Ejecta Evolution up to Seconds Timescale: Dynamics, Element Distribution, and Light Curves

We present long-term, three-dimensional simulations of ejecta from four binary neutron star mergers to second-long time-scales. Numerical-relativity data serve as boundary conditions for a general-relativistic hydrodynamics evolution incorporating an equation of state valid outside nuclear statistical equilibrium and an effective nuclear-heating prescription based on reaction-network calculations. We investigate the ejecta’s dynamical and geometrical properties, the impact of nuclear heating, the formation and spatial distribution of elements, and compute multi-angle kilonova light curves. Nuclear heating significantly affects ejecta dynamics, delaying homologous expansion beyond second time-scales and reshapes the spatial distribution of heavy nuclei. This effect is largest for asymmetric binaries with long-lived remnants; extending the evolution from ∼150 ms to ∼1 s widens the angular polar region containing 90% of the heavy-element mass (e.g., Z = 56, Z = 79) from θ ≲ 15° to θ ≲ 30°. Our nucleosynthesis results confirm that the 56Ni→56Co→56Fe decay chain dominates the heating at ∼100 days, with cobalt decay producing gamma-ray lines at 846.77 and 1238.288 keV. Comparing kilonova ray-by-ray light curves obtained from multi-angle 3D profiles and averaged 2D profiles, we find the latter approach broadly robust, although 2D light curves should be treated as upper limits. Increasing dimensionality generally lowers the bolometric luminosity, with binary asymmetry strengthening the viewing-angle dependence. For observers aligned with a lanthanide curtain’s densest region, 3D emission can match its 2D counterpart in brightness. We conclude that increased dimensionality alone is unlikely to reconcile current theoretical models with AT2017gfo observations.

Improvements to AI systems:

1. Physics-constrained generative models for ejecta morphology

• Train a generative model (e.g., normalizing flow or diffusion model) on the 3D ejecta density/composition profiles from these simulations, conditioned on binary parameters (mass ratio, remnant lifetime).

• Capability: Rapidly predict realistic 3D ejecta structures for unseen merger configurations without full GRHD simulations, enabling fast parameter-space exploration.

2. Viewing-angle-aware kilonova light-curve emulators

• Build a neural network that maps 3D ejecta profiles (or reduced features like angular mass distribution and lanthanide fraction) to multi-angle bolometric and spectral light curves, explicitly learning the asymmetry-induced viewing-angle dependence.

• Capability: Instantaneous generation of angle-dependent light curves for arbitrary ejecta profiles, replacing expensive ray-by-ray radiative transfer.

3. Nuclear-heating-aware temporal dynamics predictors

• Develop a surrogate model that predicts the delay in homologous expansion and the time-dependent angular broadening of heavy-element regions (e.g., from θ≤15° to ≤30° between 150 ms and 1 s) as a function of binary asymmetry and remnant lifetime.

• Capability: Fast forecasting of ejecta evolution and heavy-element spatial distribution for real-time analysis of merger simulations or observational follow-up.

4. Uncertainty-aware 2D-to-3D correction module

• Train a model that takes 2D averaged light curves and predicts the 3D correction factor (including upper-limit bias and viewing-angle spread) based on binary asymmetry metrics.

• Capability: Automatically adjust existing 2D kilonova models to produce more accurate 3D predictions, improving interpretation of observed events like AT2017gfo.

5. Gamma-ray line emission classifier

• Use the nucleosynthesis results (e.g., 56Ni→56Co→56Fe chain, specific gamma-ray lines at 846.77 and 1238.288 keV) to train a classifier that identifies whether a synthetic or observed light curve/spectrum is consistent with these decay signatures.

• Capability: Automated screening of kilonova candidates for cobalt decay features, aiding in distinguishing between different ejecta heating mechanisms.

6. Multi-fidelity simulation acceleration

• Implement a hierarchical AI framework that uses low-fidelity (2D, short-timescale) simulations to predict high-fidelity (3D, long-timescale) outcomes, with a learned correction for nuclear-heating and asymmetry effects.

• Capability: Reduce computational cost by orders of magnitude while retaining accuracy for heavy-element yields and light-curve predictions, enabling large-scale population studies.