RHINE Puts AI Inside Neutron Star Collisions to Unlock the Secrets of Heavy Elements

When two ultra-dense neutron stars spiral into each other and merge, the violent event sprays out neutron-rich matter that forges heavy elements like gold and platinum through the rapid neutron-capture process—or r-process. Modeling this nucleosynthesis is critical for interpreting the resulting kilonova, a transient astronomical event that gives us a direct window into cosmic element factories. But simulating the r-process during a neutron star merger has been notoriously computationally expensive, often taking weeks of supercomputing time for a single 3D model.

An international research team at GSI/FAIR has now introduced a practical solution: RHINE. Published in Physical Review D, this new framework uses a deep-learning neural network to emulate the r-process and the energy it releases, making self-consistent 3D simulations feasible in a fraction of the time .

What exactly is RHINE?

RHINE stands for R-process Heating Implementation in hydrodynamic simulations with NEural networks. It is a machine-learning framework designed to predict the nuclear heating rate and compositional changes from the r-process on the fly during a hydrodynamic simulation of a neutron star merger. Normally, a full nuclear reaction network—tracking the transformations of thousands of isotopes—would be needed at every time step in every cell of the simulation. RHINE replaces this with a small, fast neural network, cutting the computational cost dramatically .

How the deep-learning model works

RHINE employs a multilayer perceptron architecture that has been trained on thousands of reference r-process calculations from a full nuclear network. Those training data trace out the thermodynamic and compositional histories of neutron-rich matter under merger conditions. Once trained, the network takes just four locally evolved simulation quantities as input: the local density, temperature, electron fraction, and mean mass number. From these inputs, it predicts eight key source terms that govern how the r-process proceeds—including the nuclear heating rate, changes in the electron fraction, and the average atomic and mass numbers of the composition .

By injecting these predictions into the hydrodynamic simulation at each location and time step, researchers no longer need to run the full nuclear network in real time. This conceptually simple but powerful approach avoids the bottleneck that previously made long-duration or high-resolution r-process simulations impractical .

How RHINE was validated

Validation is critical whenever a machine-learning model replaces fundamental physics calculations. The RHINE team used two rigorous classes of tests to ensure the neural network was reliable under realistic conditions :

• Steady-state wind solutions: The team tested RHINE on spherically symmetric, steady-state winds covering a wide grid of parameters—ranging over electron fractions, entropies, and outflow velocities expected in merger environments. The network’s predictions matched the full network results closely across this parameter space.
• Full 3D neutron star merger simulations: As a sterner test, they applied RHINE in a complete 3D hydrodynamical simulation of a binary neutron star merger. The predictions were then compared against detailed post-processing nucleosynthesis calculated retroactively. Agreement remained high under the full complexity of a dynamic merger, indicating that the machine-learning substitute was not sacrificing physical fidelity .

Commenting on the performance, the researchers noted that the method can save a "tremendous amount of computing time" while retaining the accuracy needed for astrophysical interpretation .

Why this matters for kilonova astronomy

The energy released by the r-process directly alters the velocity, temperature, and composition of merger ejecta—all factors that shape the kilonova light curve we observe through telescopes. The landmark kilonova AT2017gfo, associated with the gravitational-wave event GW170817, gave the first detailed look at such emission, but linking that signal back to the underlying nuclear physics has been a challenge. RHINE now enables researchers to incorporate r-process heating self-consistently in 3D simulations, making it far more practical to generate theoretical predictions that can be compared directly with observed kilonovae .

Connecting to laboratory experiments at FAIR

RHINE will also serve as a computational bridge between theory and upcoming nuclear physics experiments at FAIR, the Facility for Antiproton and Ion Research in Darmstadt, Germany. FAIR will probe the properties of exotic neutron-rich nuclei that are currently beyond experimental reach but critically shape r-process outcomes. By accelerating simulations to match the speed of data analysis, RHINE offers a pathway to directly link laboratory measurements with astrophysical observations—testing models of element formation against real-world nuclear data for the first time .

Open science: Access the code

In the spirit of open science, the research team has made the RHINE source code publicly available on Zenodo, the open-access scientific repository. Researchers interested in using or building upon the method can access it at:

https://zenodo.org/records/15864447

This public release means that other simulation groups can implement RHINE in their own merger codes, expanding the impact of the framework across the broader astrophysics community.