Large-step neural network for learning the symplectic evolution from partitioned data

Xin Li, Jian Li, Zhihong Jeff Xia, Georgakarakos, NikolaosarXiv 2022 · DOI · arXiv

Abstract

In this study, we focus on learning Hamiltonian systems, which involves predicting the coordinate (q) and momentum (p) variables generated by a symplectic mapping. Based on Chen & Tao (2021), the symplectic mapping is represented by a generating function. To extend the prediction time period, we develop a new learning scheme by splitting the time series (q_i, p_i) into several partitions. We then train a large-step neural network (LSNN) to approximate the generating function between the first partition (i.e. the initial condition) and each one of the remaining partitions. This partition approach makes our LSNN effectively suppress the accumulative error when predicting the system evolution. Then we train the LSNN to learn the motions of the 2:3 resonant Kuiper belt objects for a long time period of 25000 yr. The results show that there are two significant improvements over the neural network constructed in our previous work (Li et al. 2022): (1) the conservation of the Jacobi integral, and (2) the highly accurate predictions of the orbital evolution. Overall, we propose that the designed LSNN has the potential to considerably improve predictions of the long-term evolution of more general Hamiltonian systems.

Introduction

Predicting the long-term motion of celestial bodies in Hamiltonian systems—such as planets, moons, or asteroids—presents a significant challenge in computational physics. Standard numerical methods and "black-box" neural networks often struggle to maintain the underlying symplectic geometry of phase space, leading to the accumulation of errors in position ($q$) and momentum ($p$). Over long timescales, these errors result in unphysical outcomes where fundamental constants, like energy or the Jacobi integral, are no longer conserved.

To address this, researchers have developed the Large-Step Neural Network (LSNN). This model leverages symplectic mappings via generating functions to ensure physical consistency while introducing a novel data partitioning strategy to suppress error growth over extended simulations.

Methodology

The LSNN architecture builds upon Generating Function Neural Networks (GFNN). Instead of directly predicting future coordinates, the network learns a modified generating function ($S$). This framework ensures that the transformation from time $t$ to $t+n$ is automatically symplectic, preserving volume in phase space and maintaining constants of motion.

The core innovation of the LSNN is its Data Partitioning scheme. Traditional recursive models predict the next step ($t \to t+1$) and feed that output back into the network, which compounds errors exponentially. In contrast, the LSNN trajectory is split into discrete partitions. The network is trained to map the initial condition (Partition 0) directly to multiple future windows (Partition 1, 2, 3, etc.).

The LSNN data partitioning scheme, mapping initial conditions (Partition 0) directly to multiple future time windows to prevent error accumulation.

Because generating functions define implicit relationships, the model employs Newton’s iteration to solve for final coordinates. This allows the LSNN to handle time steps ($h$) nearly 100 times larger than previous state-of-the-art models without sacrificing accuracy.

Experimental Results

The model was evaluated using the Planar Circular Restricted 3-Body Problem (PCR3BP), specifically focusing on 2:3 resonant Kuiper Belt Objects (KBOs) like Pluto. Ground truth data was generated using high-precision Runge-Kutta integrators.

Key results include:

  • Error Suppression: The LSNN demonstrated a significantly lower rate of error growth compared to standard GFNNs. By "anchoring" predictions to the initial condition, the model effectively broke the "telephone game" of error propagation common in recursive architectures.
  • Conservation of the Jacobi Integral: For resonant KBOs, the LSNN maintained the Jacobi integral with an error of less than 0.007 over a 25,000-year period.
  • Orbital Precision: The model accurately captured complex behaviors such as the libration of resonant angles, which is critical for assessing long-term orbital stability.
Comparison of coordinate error growth over time between the standard GFNN and the proposed LSNN, highlighting the effectiveness of the partitioning approach.

Conclusion

The LSNN represents a significant advancement in learning the evolution of Hamiltonian systems. By combining symplectic generating functions with a partitioned training scheme, the model achieves high-precision, long-term stability that surpasses previous neural network approaches. While the current work focuses on the 3-body problem, the partitioning strategy holds promise for scaling to more complex N-body systems, potentially enabling the study of solar system stability over million-year timescales.

Summary updated

Abstract

In this study, we focus on learning Hamiltonian systems, which involves predicting the coordinate (q) and momentum (p) variables generated by a symplectic mapping. Based on Chen & Tao (2021), the symplectic mapping is represented by a generating function. To extend the prediction time period, we develop a new learning scheme by splitting the time series (q_i, p_i) into several partitions. We then train a large-step neural network (LSNN) to approximate the generating function between the first partition (i.e. the initial condition) and each one of the remaining partitions. This partition approach makes our LSNN effectively suppress the accumulative error when predicting the system evolution. Then we train the LSNN to learn the motions of the 2:3 resonant Kuiper belt objects for a long time period of 25000 yr. The results show that there are two significant improvements over the neural network constructed in our previous work (Li et al. 2022): (1) the conservation of the Jacobi integral, and (2) the highly accurate predictions of the orbital evolution. Overall, we propose that the designed LSNN has the potential to considerably improve predictions of the long-term evolution of more general Hamiltonian systems.
BibTeX
@misc{li2022large,
  author = {Xin Li and Jian Li and Zhihong Jeff Xia and Georgakarakos, Nikolaos},
  title = {{Large-step neural network for learning the symplectic evolution from partitioned data}},
  year = {2022},
  eprint = {2208.14148},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2208.14148},
}
Figure