Data-driven discovery of non-Newtonian astronomy via learning non-Euclidean Hamiltonian

Oswin So, Gongjie Li, Evangelos A. Theodorou, Molei TaoarXiv 2022 · DOI · arXiv

Introduction

In classical celestial mechanics, planets are often simplified as point masses. However, real-world dynamics involve complex "non-Newtonian" effects such as rigid-body rotations, tidal dissipation, and general relativity, which significantly influence orbital evolution over long periods. Learning these dynamics from observational data presents two major challenges: the multiscale nature of subtle effects (like orbital precession) and the non-Euclidean geometry of rotational data, which resides on Lie group manifolds (e.g., $SO(3)$). Standard neural networks often fail to respect these geometric constraints, leading to "drift" where predicted rotation matrices lose their physical validity.

Methodology: The Lie T2 Framework

To address these challenges, researchers developed Lie T2, a recurrent neural network architecture designed to learn Hamiltonians on Lie groups. The core of this framework is the Lie T2 integrator, which possesses two critical properties:

  1. Symplecticity: It preserves phase-space volume, preventing the energy drift common in standard numerical integrators like Runge-Kutta 4 (RK4).
  2. Lie-Group Preservation: It ensures that rotation matrices remain perfectly orthogonal, keeping the system on the correct mathematical manifold.

The model employs a "splitting" technique, dividing the Hamiltonian into known components (e.g., basic kinetic energy) and unknown "residual" potentials or non-conservative forces. Multi-Layer Perceptrons (MLPs) are used to learn these unknown terms from time-series data. By embedding the Lie T2 integrator as a recurrent block, the network is constrained to obey physical and geometric laws during training.

Comparison of orbital trajectories. The Lie T2 model accurately captures the complex orbital behavior of rigid bodies, including apsidal precession, which standard point-mass models miss.

Experimental Evaluation

The method was evaluated on two primary simulations:

  • Toy Two-Body Problem: Designed to demonstrate how rigid-body corrections induce apsidal precession (the slow rotation of an orbital ellipse).
  • TRAPPIST-1 System: A complex seven-planet system simulated using the GRIT simulator, testing the model's performance with multiple interacting bodies and large time steps ($\Delta t$).

Key Insights and Results

A central finding of this work is the "Both or Nothing" Rule: long-term astronomical stability requires both symplecticity and manifold preservation. Methods that lacked one of these properties—such as Verlet (symplectic but not Lie-preserving) or Lie RK4 (Lie-preserving but not symplectic)—failed to converge when observation intervals were large.

Error analysis over integrator steps. Lie T2 maintains superior stability in both orthogonality (top) and Hamiltonian energy conservation (bottom) compared to standard baselines.

Quantitatively, Lie T2 reduced trajectory errors by up to two orders of magnitude compared to Newtonian point-mass baselines. In the TRAPPIST-1 tests, Lie T2 was the only method capable of successful training under large time steps; all other baselines diverged. Qualitatively, the model successfully "discovered" orbital precession from raw data, matching ground-truth rigid-body potentials without explicit prior knowledge of the underlying equations.

Conclusion and Future Directions

Lie T2 demonstrates that respecting the underlying geometry of physical systems is essential for data-driven discovery in astronomy. While the current model assumes known masses and inertial tensors, future work aims to learn these constants alongside dynamics. Additionally, improving robustness to noisy telescope data and scaling to larger systems via Graph Neural Networks are promising avenues for further research.

Summary updated

Abstract

Incorporating the Hamiltonian structure of physical dynamics into deep learning models provides a powerful way to improve the interpretability and prediction accuracy. While previous works are mostly limited to the Euclidean spaces, their extension to the Lie group manifold is needed when rotations form a key component of the dynamics, such as the higher-order physics beyond simple point-mass dynamics for N-body celestial interactions. Moreover, the multiscale nature of these processes presents a challenge to existing methods as a long time horizon is required. By leveraging a symplectic Lie-group manifold preserving integrator, we present a method for data-driven discovery of non-Newtonian astronomy. Preliminary results show the importance of both these properties in training stability and prediction accuracy.
BibTeX
@misc{so2022data,
  author = {Oswin So and Gongjie Li and Evangelos A. Theodorou and Molei Tao},
  title = {{Data-driven discovery of non-Newtonian astronomy via learning non-Euclidean Hamiltonian}},
  year = {2022},
  eprint = {2210.00090},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2210.00090},
}
Figure