The universe's heaviest elements arrive via cosmic violence
Heavy elements beyond iron are forged in violent cosmic events through rapid neutron capture, a process previously too complex to simulate efficiently. The new RHINE model uses deep learning neural networks to represent nuclear reaction energy, saving tremendous computing time while maintaining accuracy.
- Heavy elements beyond iron are created through rapid neutron capture during neutron star mergers and supernovae
- RHINE uses deep learning neural networks to model nuclear reaction energy in hydrodynamic simulations
- The model reduces computing time dramatically while maintaining accuracy against reference data
- Published in Physical Review D, April 2026; funded partly by the European Research Council
Researchers developed RHINE, a machine learning model that simulates how heavy elements like gold and uranium are created during neutron star mergers and supernovae, dramatically reducing computational demands.
Gold arrives on Earth from the cosmos. So does uranium, platinum, and nearly every element heavier than iron. For decades, scientists have known these materials are forged in the universe's most violent moments—the collision of neutron stars, the death throes of massive suns—but understanding exactly how has remained stubbornly out of reach. The physics is real. The problem is computational: simulating these events in full detail demands so much computing power that researchers have had to strip away crucial details just to make the math tractable.
A team at GSI/FAIR, a research institution in Germany, along with international collaborators, has found a way around this bottleneck. They built a machine learning model called RHINE that dramatically reduces the computational burden while preserving the accuracy scientists need. The breakthrough, published in Physical Review D in April 2026, represents the first time researchers have embedded a deep learning neural network directly into hydrodynamic simulations of element formation during extreme cosmic events.
The physics at work is called the r-process, shorthand for rapid neutron capture. When neutron stars collide or a massive star explodes as a supernova, the environment becomes flooded with free neutrons and unimaginable energy. Atomic nuclei in this maelstrom absorb neutrons at extraordinary speed. Those neutrons decay into protons, building heavier and heavier elements in a cascade that happens in seconds. This process is responsible for creating most of the heavy elements we observe in the universe. But modeling it requires tracking thousands of nuclear reactions simultaneously, each with its own energy output, each influencing how material moves and radiates.
Dr. Oliver Just, the lead researcher on the project, explained the core problem plainly: full simulations demand computing resources so vast that researchers have historically had to simplify their models, cutting away details that might matter. RHINE offers an alternative. The team trained a deep learning neural network using a large set of reference calculations that included the complete set of nuclear reactions. Once trained, the network could approximate the energy released during the r-process—the "heating," in technical terms—with minimal computational cost. This heating matters enormously. It influences how fast material gets ejected from the collision, how that material spreads, and what electromagnetic signals the event produces, including the kilonovae that astronomers observe after neutron star mergers.
Dr. Zewei Xiong, who designed the machine learning models, described the validation process. The team compared RHINE's output against their reference data in detail. The agreement was strong enough to confirm that machine learning could save tremendous computing time without sacrificing accuracy. More importantly, the results revealed that r-process heating is a significant effect that previous simplified models had underestimated or ignored entirely—a finding that will reshape how future simulations are built.
The implications extend beyond pure astrophysics. The FAIR facility, a major experimental complex under construction, will soon generate data about nuclear reactions under extreme conditions. RHINE could serve as a bridge, connecting what experimentalists measure in the lab with what astronomers observe in the sky. By running faster, more detailed simulations, researchers can test whether their understanding of element formation is correct. The code is publicly available, funded in part by the European Research Council, meaning other teams can adopt and refine it. The next generation of cosmic collision simulations is already being written.
Notable Quotes
Our new model RHINE, which uses artificial intelligence, offers an efficient alternative to simplified models that require incredible computing power.— Dr. Oliver Just, GSI/FAIR
The high degree of agreement suggests that machine learning models can save a tremendous amount of computing time while maintaining accuracy.— Dr. Zewei Xiong, GSI/FAIR
The Hearth Conversation Another angle on the story
Why does it matter where gold comes from? Isn't that just curiosity?
It's deeper than that. Understanding where heavy elements form tells us about the history of the universe itself. Every gold atom in existence came from somewhere specific—a neutron star collision, most likely. If we can't simulate those events accurately, we're missing a piece of how the cosmos works.
But you said the simulations were already possible. What changed?
They were possible, but crippled. Imagine trying to paint a portrait with only three colors because you don't have time to mix more. You get the outline, maybe, but you lose the face. RHINE lets us use the full palette without waiting forever for the computer to finish.
How does a neural network know anything about nuclear physics?
It doesn't, not inherently. The network is trained on thousands of detailed calculations that physicists already did the hard way. It learns the patterns—given these conditions, this much energy gets released. Then it can predict new cases instantly instead of recalculating from scratch.
So it's a shortcut.
A very sophisticated shortcut. But here's what matters: when they tested it against the real data, it didn't cut corners. The accuracy held up. That's rare.
What happens next?
The FAIR facility will start producing experimental data about nuclear reactions under extreme pressure. RHINE can now connect that lab data to what we see in the sky—to actual neutron star collisions. We'll finally be able to test whether our theories are right.