When Neutron Stars Collide, the Aftermath Is an Inferno. Now AI Can Simulate It.
24 Apr 2026, Yanjiang
AI now simulates the fiery aftermath of neutron star collisions, revealing how extreme heat drives element formation.
What happens in the milliseconds after two neutron stars crash into each other? The answer, as far as any human will ever directly witness, is something close to hell. Temperatures reach a trillion degrees. Densities exceed that of atomic nuclei. And in that inferno, the universe assembles its heaviest elements — gold, platinum, uranium — through a process so complex that even our most powerful supercomputers struggle to track it in real time.
This is the r-process: rapid neutron capture, the cosmic forge that builds nuclei heavier than iron. For decades, astrophysicists have known it happens in neutron-star mergers. They have also known that the energy released by this nuclear alchemy — r-process heating — can dramatically alter the dynamics of the explosion itself. Faster ejecta, brighter kilonovae, different element yields. The problem is that a full nuclear network, tracking thousands of isotopes through tens of thousands of reactions, is computationally prohibitive inside a hydrodynamic simulation.
So most simulations cheat. They either ignore r-process heating entirely, or approximate it with crude parametrizations. It is like modeling a forest fire by assuming every tree burns at the same rate — technically possible, but wrong in every interesting way.
Now, a team led by Gabriel Martínez-Pinedo at GSI Helmholtzzentrum in Germany — working with Oliver Just and Zewei Xiong — has proposed a radically different solution. Their method, called RHINE, uses machine learning to emulate the r-process and its energy release inside hydrodynamic simulations. The preprint (arXiv:2507.09040) describes a framework that achieves better than 10% accuracy in heating energy while requiring the evolution of only a handful of additional quantities — a fraction of the computational cost of a full nuclear network.
To understand why this matters, we first need to look at what happens when neutron stars merge.
The cosmic kitchen
Neutron stars are the collapsed cores of massive stars — objects the mass of the Sun compressed into a sphere the size of a city. When two of them spiral together and collide, they eject material under extreme conditions: neutron-rich, hot, and expanding rapidly.
In this ejected material, neutrons are captured by seed nuclei faster than those nuclei can beta-decay. The nuclei climb the chart of isotopes, absorbing neutrons one after another, until they become so neutron-rich that they can no longer hold together. Then they fission, releasing more neutrons, and the process continues. This is the r-process, and it is responsible for about half of all elements heavier than iron.
The energy released by these nuclear reactions — about 2 to 3 MeV per baryon in the simulations described in the paper — does not just dissipate. It heats the surrounding material, increasing its pressure and accelerating it outward. This is r-process heating, and it can change the velocity distribution of the ejecta by tens of percent. In some cases, as the team found for black-hole torus ejecta, the heating makes the material 40% more massive — because slower-moving material that would otherwise fall back onto the remnant gets pushed out instead.
This is not a small effect. It changes the shape of the explosion, the brightness of the kilonova, and potentially the yields of the elements produced. But until now, capturing it in simulations required either heroic computational effort or crude approximations.
The neural forge
RHINE stands for “R-process Heating In Neural-network Emulation,” and its design is elegant in its simplicity. Instead of tracking thousands of isotopes, the method evolves just four mass fractions — neutrons, protons, alpha particles, and a lumped “heavy nuclei” category — plus the average mass number of heavy nuclei and the average mass excess per baryon. That is six quantities instead of thousands.
The trick is that the rates of change for these quantities — the source terms that determine how fast the composition evolves and how much energy is released — are predicted by neural networks. These networks were trained on a large set of trajectories from full nuclear-network calculations, covering a wide range of thermodynamic conditions: temperatures from billions of degrees down to a hundred million, densities from neutron-star crust to near-vacuum, electron fractions from neutron-rich to proton-rich.
The architecture is a standard multilayer perceptron — nothing exotic, no transformer layers or attention mechanisms. Just layers of perceptrons, each taking input from the previous layer and passing output to the next. The team trained separate networks for different thermodynamic regimes: one set for when the material is in nuclear statistical equilibrium (above 7 GK), another for quasi-statistical equilibrium (between 5 and 7 GK), and a third for the dynamic regime where no equilibrium conditions apply. This is not a single black box; it is a carefully orchestrated ensemble of specialized predictors, each trained for its specific domain.
The result is a method that can be plugged into any existing hydrodynamics code with minimal modification. The team provides pre-trained machine-learning data and routines for source-term prediction online. Other researchers can download them and start simulating immediately.
What the simulations reveal
The team tested RHINE on two types of simulations: spherically symmetric wind models, which provide a controlled testbed, and full three-dimensional neutron-star merger simulations, which capture the real complexity of the event.
In the wind models, the comparison between RHINE and full nuclear-network post-processing showed remarkable agreement. The heating energy matched to within 10% across a range of initial conditions — different electron fractions, different entropies, different expansion timescales. The velocity boosts, the composition evolution, the final element distributions — all were captured faithfully.
The merger simulations told a more dramatic story. Without r-process heating, the ejecta from a neutron-star merger expands in a roughly homologous fashion: faster material outruns slower material, and the velocity profile is set early. With r-process heating, everything changes. The heating accelerates the ejecta, especially the slower-moving material from the black-hole torus. This material inflates and becomes more spherical, smoothing out small-scale variations in the density distribution.
The team found that about 2.3 MeV per baryon is released in dynamical ejecta, 0.7 MeV in neutron-star torus ejecta, and 2.1 MeV in black-hole torus ejecta. The strongest velocity boost occurs in the black-hole torus ejecta, which also become 40% more massive — meaning more material escapes the gravitational pull of the remnant.
The nucleosynthesis yields are only mildly affected by r-process heating — the same elements are produced, in roughly the same proportions. But the kilonova signal changes significantly. Once the black-hole torus ejecta become visible, the kilonova gets substantially brighter. This is not a subtle effect; it could determine whether a given merger is detectable by telescopes like the Vera Rubin Observatory or the James Webb Space Telescope.
The limits of the analogy
Like an investment portfolio where safe bets fund high-risk ventures, RHINE balances computational efficiency against accuracy — but unlike portfolio management, where risk can be diversified arbitrarily, the neural network’s predictions are only as good as its training data. The team trained on nuclear-network calculations that themselves rely on nuclear physics inputs with significant uncertainties — reaction rates, masses, fission yields. RHINE inherits these uncertainties, even as it reduces the computational cost of incorporating them into simulations.
This is not a weakness of the method, but a reminder that machine learning cannot transcend the quality of its training data. The neural networks learn the patterns present in the nuclear-network calculations; they do not discover new physics.
The philosophical implications
What RHINE represents is something deeper than a technical advance. It is a shift in how we approach computational astrophysics — from brute force to intelligent approximation, from tracking every detail to understanding which details matter.
For decades, the standard approach to simulating neutron-star mergers has been to either ignore r-process heating or treat it with crude parametrizations. The implicit assumption was that the heating was too complex to model accurately within a hydrodynamic simulation, so any approximation was better than nothing. RHINE challenges this assumption by demonstrating that accurate emulation is possible with minimal computational overhead.
This is not a question from a philosophy seminar. It is the practical reality of modern computational science: we are no longer limited by whether we can simulate something, but by whether we can simulate it efficiently enough to explore the parameter space. RHINE opens up that parameter space for r-process heating, allowing simulations that were previously impossible.
The team is already working on extending the method to other explosive environments — core-collapse supernovae, accretion disks around black holes. The same approach could be applied to any problem where a complex microphysical process needs to be emulated within a larger simulation: neutrino transport, magnetic field evolution, chemical reaction networks.
Perhaps one day, when astrophysicists run simulations of neutron-star mergers, they will take accurate r-process heating for granted — a standard component of the code, like gravity or hydrodynamics. When that day comes, they will have methods like RHINE to thank. Not for discovering new physics, but for making the old physics computable.
Yanjiang is an online editor of Loom Science
References
- O. Just et al., R-process heating implementation in hydrodynamic simulations with neural networks, arXiv:2507.09040