Machine Learning Speeds Up Neutron Star Merger Simulations
Researchers at GSI/FAIR have built a machine-learning system that estimates the heat released during neutron star collisions, solving a long-standing computational bottleneck in astrophysical modeling. The system, called RHINE, tracks energy from rapid neutron-capture nucleosynthesis-the r-process-while hydrodynamic simulations run, making it possible to include effects that were previously too expensive to calculate.
The r-process is how the universe builds heavy elements like gold and platinum in extreme environments. When neutron stars collide or massive stars explode, free neutrons are captured by existing nuclei, forming heavier atoms. Modeling this process exactly requires tracking thousands of isotopes simultaneously, which creates a computational wall in multidimensional simulations.
RHINE sidesteps that wall by following a much smaller set of quantities instead. Neural networks trained on detailed nuclear-reaction calculations then estimate the energy released during the r-process. The approach is the first to embed machine-learning models directly into multidimensional hydrodynamical simulations of this kind.
Why the Heat Matters More Than It Appears
The heat generated during the r-process is not a minor detail. Several megaelectronvolts of energy per baryon can be released, and that energy changes how fast material moves and spreads through space.
The effect is most dramatic in slow-moving ejecta. In wind tests, about 3 megaelectronvolts per baryon had little effect on material already moving at 0.3 times the speed of light. But the same heat nearly tripled the final speed of slower material that would have reached only 0.03 times the speed of light. Basic energy conservation explains the pattern: slower material is easier to accelerate.
Many existing simulations use postprocessing, computing fluid motion first and adding nuclear reactions later along extracted trajectories. That saves time but leaves the fluid unable to "feel" the extra heat while moving. RHINE fixes this by calculating heating rates during the simulation itself.
Effects on Merger Predictions
In full neutron star merger simulations, RHINE produced measurable changes. The slow black-hole-torus ejecta-matter expelled later from the black hole and surrounding torus-gained about 2.1 megaelectronvolts per baryon of heating energy. Its average velocity increased by roughly 40 percent, and its mass rose from 0.04929 to 0.06 solar masses.
The faster dynamical ejecta behaved differently. They received slightly more heating energy but showed only modest velocity changes because they were already moving much faster. The overall abundance pattern of heavy elements stayed largely the same, though individual nuclei shifted by factors of a few in some cases.
The biggest impact appeared in kilonova brightness-the electromagnetic glow produced by freshly forged heavy elements after a merger. Including r-process heating made the kilonova roughly twice as bright, largely because slow ejecta became both more massive and faster.
A Tool for Routine Use
The researchers validated RHINE against full nuclear-network calculations in both idealized models and long-term merger simulations. In most cases, the net heating energy agreed within about 10 percent. The code is publicly available.
RHINE is useful not because it overturns the big picture of neutron star mergers, but because it improves the details without crushing computational costs. The scheme adds only a modest computational burden and avoids extra tracer-particle machinery.
The method points to a broader possibility: if machine learning can stand in for parts of a nuclear reaction network without losing crucial physics, similar approaches might work for other hard-to-compute ingredients in astrophysical simulations. The researchers say RHINE could eventually help connect future experiments at FAIR with observations of stellar explosions and neutron star mergers.
The findings appear in Physical Review D. The project was co-funded in part by the European Research Council.
Your membership also unlocks:
