Data-Driven Continuum Dynamics via Transport-Teleport Duality

Jong-Hoon AhnarXiv 2020 · arXiv

Introduction

Modeling the dynamics of continuous media, such as fluids or moving objects, typically requires dense temporal data and complex governing equations. This paper introduces a novel mathematical framework—the Transport-Teleport Duality—designed to overcome the limitations of traditional deep learning in physics. By re-representing physical transport as a velocity-free "teleportation" process, the proposed method enables the learning of complex dynamics from extremely sparse observations, such as just two or three frames, without prior knowledge of specific physical laws.

The Challenge of Physical Representation

Deep learning models in physics often struggle with high data requirements, lack of interpretability, and physical inconsistencies (e.g., violating conservation of mass). These issues often stem from the standard representation of physical data, particularly the reliance on the velocity field. Tracking mass movement through space-time is computationally intensive. By shifting the representation, we can simplify the underlying learning problem.

Methods: The Inversion Transform

The core of this approach is an "inversion transform" that maps physical transport phenomena into a dual "teleportation" space.

The Transport-Teleport workflow: Inverting the Eulerian view to enable point-wise interpolation and prediction.

Transport-Teleport Duality

The transform maps mass density ($\rho$) to its reciprocal ($\tilde{\rho} = 1/\rho$) and negates local rotation angles. In this inverted space, the effective velocity field becomes zero. Instead of mass transporting from one location to another, the quantity appears to "teleport"—disappearing at one point and reappearing at another. This converts a complex spatio-temporal problem into a simpler point-wise process.

Optimization and Interpolation

Dynamics are recovered by solving an optimization problem that minimizes the change in inverted mass density between observations. Once the inverted representation is established, standard second-order Lagrange polynomials are used for interpolation or extrapolation. These results are then transformed back into the original physical space to produce smooth, continuous motion.

Experimental Results

The method was evaluated using sparse datasets, including a soccer ball in flight (3 frames), a running cheetah (2 frames), and deforming handwritten digits (2 frames).

Reconstructing a soccer ball's trajectory from three sparse observational frames using the inversion transform.

Unlike traditional deep learning, which requires thousands of training examples to generalize, this method learns specific trajectories directly from a handful of frames. For the soccer ball example, the inversion transform successfully generated a physically plausible curved trajectory. While standard pixel-wise interpolation results in "ghosting" or blurry overlaps, this dual representation maintains the integrity of the moving object.

Key Insights

  • Velocity-Free Dynamics: Removing velocity from the representation eliminates the primary source of complexity in continuum mechanics, allowing simple models to outperform deep architectures.
  • Inherent Conservation: The transform is rooted in the Jacobian of the coordinate mapping, ensuring that the resulting flow inherently satisfies the continuity equation. Mass conservation is a property of the math, not a learned constraint.
  • Efficiency: The approach reduces model complexity significantly, requiring neither dozens of neural layers nor massive "black box" training sets.

Limitations and Future Outlook

While highly effective, the method faces challenges with 2D natural images where occlusions (e.g., a cheetah's leg moving behind its body) violate mass conservation. The author suggests that the framework is most naturally suited for 3D volumetric data, such as fluid simulations or medical imaging, where all mass is accounted for. Future work involves integrating this representation as a pre-processing step for AI models to explore "new physics" with minimal data.

Summary updated

Abstract

In recent years, machine learning methods have been widely used to study physical systems that are challenging to solve with governing equations. Physicists and engineers are framing the data-driven paradigm as an alternative approach to physical sciences. In this paradigm change, the deep learning approach is playing a pivotal role. However, most learning architectures do not inherently incorporate conservation laws in the form of continuity equations, and they require dense data to learn the dynamics of conserved quantities. In this study, we introduce a clever mathematical transform to represent the classical dynamics as a point-wise process of disappearance and reappearance of a quantity, which dramatically reduces model complexity and training data for machine learning of transport phenomena. We demonstrate that just a few observational data and a simple learning model can be enough to learn the dynamics of real-world objects. The approach does not require the explicit use of governing equations and only depends on observation data. Because the continuity equation is a general equation that any conserved quantity should obey, the applicability may range from physical to social and medical sciences or any field where data are conserved quantities.
BibTeX
@misc{ahn2020data,
  author = {Jong-Hoon Ahn},
  title = {{Data-Driven Continuum Dynamics via Transport-Teleport Duality}},
  year = {2020},
  eprint = {2005.13358},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2005.13358},
}
Figure