Proefschrift
Solving the Gravitational N-body Problem with Machine Learning
In this work, I explore the creation of new methods that optimize simulations of the gravitational N-body problem. Specifically, I take advantage of the recent popularity of Machine Learning methods to find tools that can suit this problem.
- Auteur
- V. Saz Ulibarrena
- Datum
- 07 oktober 2025
- Links
- Thesis in Leiden Repository
Firstly, running simulations with large numbers of bodies can become extremely computationally expensive. By replacing these expensive parts with neural networks, the total time to run a simulation can be reduced by orders of magnitude. We take advantage of a specific type of neural networks that include physical knowledge into their structure: physics-aware neural networks. Secondly, setting up a simulation requires the selection of multiple simulation parameters. In most cases, expert knowledge is required to make the right choice for each specific system. A wrong choice can lead to the simulation being incorrect, and therefore unusable for scientific discovery. I create a reinforcement learning method that automatically chooses the time-step size. With my method, this value is chosen to balance accuracy and speed in the simulation, leading to optimum results without the need of expert knowledge. My work presents an initial exploration of the capabilities of Machine Learning techniques for a complex problem: the simulation of multiple bodies moving under their mutual gravity. I identify the challenges of this problem and present innovative solutions.