🔭 astrophysics

Terrestrial planet formation in the era of GPU computing

This paper introduces the GPU-accelerated GENGA code to unify symplectic integrators for long-term orbital and short-term collisional dynamics, demonstrating through comparative simulations that avoiding acceleration factors prevents chemical distortions, that formation timescales depend on initial planetesimal sizes, and that terrestrial planets can form resonant chains without gas-driven migration.

Original authors: Simon L. Grimm, Joachim G. Stadel

Published 2026-04-13

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Simon L. Grimm, Joachim G. Stadel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to predict how a messy pile of LEGO bricks will eventually snap together to form a castle. Now, imagine that pile contains 65,000 bricks, they are floating in a giant vacuum, and you need to watch them collide and merge over billions of years.

That is essentially what this paper is about. It's a guide to how scientists are using super-fast computer chips (called GPUs, the same ones gamers use for video games) to simulate how rocky planets like Earth are born.

Here is the story of the paper, broken down into simple concepts and analogies.

1. The Old Way vs. The New Way (The Hardware Revolution)

For a long time, simulating planet formation was like trying to count every grain of sand on a beach using a single abacus. It took forever. Scientists had to cheat to make it work. They used a "speed-up button" (called an acceleration factor) that made the planets grow faster than they should in real life. It was like fast-forwarding a movie so much that the actors started running around the room.

The New Era:
The authors, Simon Grimm and Joachim Stadel, built a new tool called GENGA. Think of GENGA as a massive, high-speed assembly line powered by thousands of tiny workers (the GPU cores) working in perfect unison.

The Analogy: If an old computer (CPU) is a single chef chopping vegetables, a GPU is a stadium full of 10,000 chefs chopping at the same time.
The Result: Because these computers are so fast, scientists no longer need to hit the "speed-up button." They can run simulations at real speed. This is crucial because, as the paper shows, hitting the speed-up button actually changes the final recipe of the planet, making it taste different (chemically) than it should.

2. The "Hybrid" Dance (The Math Problem)

Simulating gravity is tricky. Usually, planets orbit the sun in a smooth, predictable dance (like a waltz). But sometimes, two planets get too close and start a chaotic, violent tango.

The Problem: A standard computer program tries to calculate every single step of every dance, which is slow.
The Solution (Hybrid Symplectic Integrator): The authors use a "smart switch."
- When planets are far apart, the computer uses a fast, smooth "waltz" algorithm (Keplerian orbits).
- When planets get close and start crashing or swinging wildly, the computer instantly switches to a "slow-motion, high-detail" mode to calculate the crash accurately.
- The Metaphor: Imagine driving a car. On the highway, you cruise at 70 mph (fast mode). But when you see a deer jump out, you slam on the brakes and switch to a slow, careful crawl (detailed mode). GENGA does this automatically and instantly for thousands of planets at once.

3. The Big Discovery: Don't Cheat! (The Acceleration Factor)

The paper tested what happens if you do use that old "speed-up button" (making planets grow faster by artificially making them bigger).

The Finding: It turns out, cheating ruins the flavor.
The Analogy: Imagine baking a cake. If you rush the process by turning the oven to maximum heat, the cake might rise, but the inside will be raw, and the ingredients won't mix properly.
The Science: When they sped up the simulation, the planets formed from a very narrow ring of material (very local). When they ran it at real speed, the planets ate up material from a much wider area. This means the chemical makeup of the planet (what it's made of inside) is totally different. If we want to know if an alien planet could support life, we can't use the "speed-up" method; we need the real-time simulation.

4. Resolution Matters: Pixelated vs. HD

The authors also asked: "How many tiny rocks (planetesimals) do we need to start with to get a good result?"

Low Resolution (2,000 rocks): Like watching a movie in 144p. It looks okay, but you miss the details. The planets formed quickly, but the collisions were violent, like two cars smashing into each other.
High Resolution (65,000 rocks): Like watching in 4K Ultra HD.
- The Surprise: With more rocks, the planets took twice as long to form.
- The Collisions: Instead of violent car crashes, the planets grew by gently absorbing tiny pebbles (like a snowball rolling down a hill).
- The Outcome: The planets in the high-res simulation ended up with a more "mixed" diet of ingredients and were less likely to be flung out of the solar system. They also naturally formed a "train" of planets locked in a rhythmic dance (resonance chains) without needing gas to push them there.

5. The Gas Giants' Role

Finally, they added Jupiter and Saturn to the mix.

The Chaos: Gas giants are like bullies in a playground. They shake things up. In these simulations, the results were chaotic—sometimes you get 3 planets, sometimes 2. It's impossible to predict exactly what will happen in one single run.
The Asteroid Belt: However, the high-resolution simulation was the only one detailed enough to keep a tiny "asteroid belt" alive. This is important because that belt acts as a delivery truck, bringing water and volatiles to the planets later on.

The Bottom Line

This paper is a celebration of computing power meeting scientific curiosity.

By using the massive power of modern graphics cards, scientists can finally stop "cheating" with speed-up buttons and run simulations that look at the universe in High Definition. They've learned that:

Don't rush the process: Real-time simulations give us the correct chemical recipe for planets.
More detail is better: Simulating more tiny rocks changes how long it takes to form a planet and how violent the collisions are.
Nature is efficient: Planets can form beautiful, rhythmic chains just by gravity, without needing gas to push them around.

The authors conclude that we are just getting started. With even more powerful computers, they hope to simulate one million interacting particles to finally solve the mystery of how our own Solar System, and the thousands of others, came to be.

1. Problem Statement

The formation of terrestrial planets involves complex gravitational interactions ( $N$ -body dynamics) over billions of years. Traditional simulations face two primary challenges:

Computational Cost: Direct $N$ -body integration has a computational complexity of $O(N^2)$ . Simulating realistic systems with high particle counts (to represent small planetesimals) was historically prohibitively expensive, often requiring billions of CPU hours.
Numerical Artifacts: To circumvent computational limits, previous studies often employed an acceleration factor ( $f$ ), artificially increasing the physical radii of bodies to speed up collision rates. This reduced the growth timescale by $f^{-2}$ but introduced unknown physical biases.
Resolution Limitations: Low-resolution simulations (fewer particles) fail to capture fine-grained dynamical processes, such as the formation of asteroid belts, detailed chemical mixing, and the specific nature of impact events (giant impacts vs. smooth accretion).

The authors aim to leverage modern Graphics Processing Units (GPUs) and advanced hybrid symplectic integrators to run high-resolution, long-term simulations without artificial acceleration factors, thereby revealing more accurate physical insights into planetary assembly.

2. Methodology

A. Numerical Integration: Hybrid Symplectic Integrators

The paper utilizes the GENGA code, which implements a hybrid symplectic integrator. This method addresses the dual challenge of long-term orbital stability and short-term close encounters:

Hamiltonian Splitting: The system's Hamiltonian is split into three parts:
1. $H_A$ : Keplerian motion around the central star (solved analytically).
2. $H_B$ : Mutual gravitational interactions between bodies.
3. $H_C$ : Center-of-mass momentum correction (due to mixed coordinates).
Close Encounters: When bodies approach within a critical radius ( $r_{crit}$ ), the integrator switches from the symplectic method to a direct, non-symplectic solver with smaller time steps.
Multi-level Switching: To handle systems where many bodies are in close proximity (common in dense disks), GENGA uses a multi-level transition. Instead of a binary switch, it reduces time steps hierarchically for affected bodies before engaging the direct solver, minimizing integration errors.

B. Hardware Architecture

GPU Acceleration: The code is optimized for NVIDIA (CUDA) and AMD (HIP) GPUs, exploiting massive parallelism for the $O(N^2)$ force calculations.
Hybrid CPU/GPU Strategy: The code employs self-tuning routines. For large $N$ (thousands to tens of thousands), it utilizes GPUs. For small $N$ (where GPU overhead dominates), it automatically switches to parallelized CPU execution (OpenMP/C++).
Additional Physics: The code incorporates non-gravitational effects such as gas drag, General Relativity (post-Newtonian corrections), tidal forces, and the Yarkovsky effect, though the primary focus of the presented results is on pure gravitational dynamics.

C. Simulation Setup

The authors conducted comparative studies using GENGA II with the following parameters:

System: A disk of planetesimals between 0.5 AU and 4 AU with a total mass of $5 M_{\oplus}$ (Earth masses).
Resolution: Simulations ranging from $N=2,048$ to $N=65,536$ initial bodies.
Variables:
- Acceleration Factor ( $f$ ): Comparing $f=1$ (physical) vs. $f=6$ (accelerated).
- Gas Giants: Simulations run both with and without Jupiter/Saturn analogs.
- Timescales: Simulations evolved over 4 billion years (Gyr).

3. Key Contributions

Demonstration of GPU Viability: The paper proves that modern GPUs can accelerate $N$ -body simulations by a factor of 150 compared to single-core CPUs and 20 compared to 16-core CPUs for large $N$ , making high-resolution, billion-year simulations feasible.
Debunking the Acceleration Factor ( $f$ ): The authors provide rigorous evidence that the artificial acceleration factor $f$ distorts the physical assembly of planets.
Resolution-Dependent Dynamics: The study quantifies how increasing particle resolution changes formation timescales, impact violence, and material mixing, showing that low-resolution simulations are insufficient for studying planetary interior composition.
Resonant Chain Formation: The simulations demonstrate that terrestrial planets can naturally form resonant chains without requiring gas-driven orbital migration, provided the system is sufficiently resolved.

4. Key Results

A. Impact of the Acceleration Factor ( $f$ )

Distorted Composition: Simulations with $f=6$ resulted in planets formed from very local material (narrow feeding zones). In contrast, $f=1$ simulations showed planets accreting from a much broader range of semi-major axes, leading to more chemically mixed interiors.
Orbital Architecture: The $f=6$ simulations produced more compact systems with weaker resonances (e.g., 3:2, 5:3) compared to the strong 2:1 resonances seen in $f=1$ runs.
Conclusion: The use of $f$ should be avoided in modern studies, especially when investigating habitability or chemical composition, as it fundamentally alters the accretion history.

B. Particle Resolution Effects

Formation Timescale: High-resolution simulations ( $N=65,536$ $N = 65, 536$ ) took approximately twice as long to form planetary cores compared to low-resolution runs ( $N=2,048$ $N = 2, 048$ ).
- Example: Forming a $0.5 M_{\oplus}$ body took ~50 Myr in low-res vs. ~90 Myr in high-res (without gas giants).
Impact Violence: Low-resolution runs were dominated by fewer, more violent "giant impacts." High-resolution runs showed smoother growth tracks with accretion dominated by smaller bodies, implying different internal structures and atmospheric compositions for the final planets.
Ejected Mass: In simulations with gas giants, higher resolution led to less ejected mass (0.5 $M_{\oplus}$ ejected in high-res vs. 1.0 $M_{\oplus}$ in low-res). This suggests low-resolution models overestimate the dynamical instability of the system.

C. Gas Giants and Resonances

In systems with gas giants, the formation process is more chaotic. While the final number of planets varied between individual runs (due to chaos), high-resolution runs were the only ones to successfully preserve a remnant asteroid belt between resonances, allowing for late-stage volatile delivery.
Terrestrial planets formed resonant chains naturally in both gas-free and gas-rich environments, challenging the necessity of gas-driven migration for resonance formation.

5. Significance and Future Outlook

Scientific Impact: The study shifts the paradigm of terrestrial planet formation modeling. It establishes that high-resolution, unaccelerated simulations are now computationally feasible and necessary to accurately model planetary composition, habitability, and formation timescales.
Methodological Advancement: The development of GENGA, with its hybrid CPU/GPU architecture and multi-level symplectic switching, sets a new standard for $N$ -body codes.
Future Directions:
- Scale: The authors propose a target of 1 million interactive particles to fully resolve the planetesimal disk and collisional fragmentation.
- Algorithms: Implementation of the Fast Multipole Method (FMM) to reduce complexity from $O(N^2)$ to $O(N)$ for even larger systems.
- Physics: Integration of detailed collision models (fragmentation) and realistic gas disk hydrodynamics to link simulation outcomes with observational constraints on exoplanet chemistry.

In summary, this paper demonstrates that the era of GPU computing has resolved the bottlenecks of terrestrial planet formation simulations, revealing that previous low-resolution, accelerated models significantly misrepresented the chemical and dynamical history of planetary systems.