Pioneer: Physics-informed Riemannian Graph ODE for Entropy-increasing Dynamics

Li Sun, Ziheng Zhang, Zixi Wang, Yujie Wang, Qiqi Wan, Hao Li, Hao Peng, Philip S. Yu

Introduction

Modeling the evolution of complex, interacting systems—such as global weather patterns or dynamic social networks—is a fundamental challenge in data science. While Graph Ordinary Differential Equations (Graph ODEs) have recently emerged as a powerful paradigm for capturing continuous-time dynamics, they typically operate in Euclidean space. This "flat-world" assumption often leads to "geometry-unaware" trajectories that fail to respect the intrinsic non-Euclidean structure of real-world data (e.g., the Earth's spherical surface). Furthermore, standard models frequently ignore the second law of thermodynamics, potentially generating trajectories that are physically impossible due to decreasing entropy.

To address these limitations, we present Pioneer, a physics-informed Riemannian graph ODE framework. Pioneer integrates Riemannian geometry with thermodynamic principles to ensure that predicted trajectories remain on the appropriate manifold and align with the physical reality of increasing system disorder.

The Pioneer Framework

Pioneer introduces three core innovations to bridge the gap between geometric deep learning and physical laws:

  1. Manifold ODE: Pioneer defines its differential system directly on a Riemannian manifold using the $\kappa$-stereographic model. This ensures that latent representations of system states are constrained to the correct geometric surface throughout their evolution.
  2. Constrained Ricci Flow: To model the structural evolution of the system (i.e., changing relationships between entities), Pioneer employs a novel version of Ricci flow. This flow is governed by a neural constraint that provably ensures non-decreasing von Neumann entropy, mirroring the thermodynamic behavior of real-world systems.
  3. Gyro-transform: To improve computational efficiency, Pioneer replaces expensive exponential and logarithmic maps with the Gyro-transform. This manifold-preserving operation utilizes simple matrix-vector multiplications, reducing numerical errors and overhead compared to traditional hyperbolic or spherical trigonometric functions.
The Pioneer architecture consists of a Riemannian Encoder, a Physics-informed Differential System governed by Constrained Ricci Flow, and a Riemannian Decoder utilizing Gyro-transforms.

Experimental Evaluation

The framework was evaluated on three diverse real-world datasets:

  • Weather: Meteorological data where the underlying geometry is the Earth's sphere.
  • CReSIS: Glacier thickness measurements in Greenland, characterized by complex spatial-temporal dynamics on curved surfaces.
  • Social: Time-evolving social network interactions.

Pioneer was compared against nine strong baselines, including state-of-the-art graph ODEs like CG-ODE and HOPE.

Results and Discussion

Pioneer consistently outperformed all baselines across various metrics. On the Social dataset, it achieved a Mean Absolute Percentage Error (MAPE) of 9.15%, a significant improvement over the next best model, HOPE (10.71%).

A key finding is the framework's robustness as prediction length increases. As shown in the figure below, the Riemannian-based modeling in Pioneer maintains lower error rates compared to Euclidean counterparts, particularly in scenarios where geometric constraints are critical for long-term stability.

Comparison of MAPE for Riemannian vs. Euclidean modeling on Social and Weather datasets across different prediction lengths.

Beyond accuracy, Pioneer successfully captures the entropy-increasing trends observed in real-world systems. Visualizations of generated states confirm that Pioneer's trajectories follow physical laws, whereas previous models often produced thermodynamically inconsistent results where system disorder decreased over time.

Conclusion

Pioneer represents a significant step forward in dynamic system modeling by enforcing both geometric and physical constraints. By operating on Riemannian manifolds and incorporating entropy-increasing Ricci flow, it provides more accurate and physically plausible predictions than traditional Euclidean models. Future work will explore extending the framework to manifolds with non-constant curvature and incorporating additional physical laws such as energy conservation.

Summary updated

BibTeX
@inproceedings{sun2025pioneer,
  author = {Li Sun and Ziheng Zhang and Zixi Wang and Yujie Wang and Qiqi Wan and Hao Li and Hao Peng and Philip S. Yu},
  title = {{Pioneer: Physics-informed Riemannian Graph ODE for Entropy-increasing Dynamics}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i12.33372},
}

Loading related content...

Figure