Modeling Randomly Observed Spatiotemporal Dynamical Systems

Valerii Iakovlev, Harri Lähdesmäki

Abstract

Spatiotemporal dynamics models are fundamental for various domains, from heat propagation in materials to oceanic and atmospheric flows. However, currently available neural network-based spatiotemporal modeling approaches fall short when faced with data that is collected randomly over time and space, as is often the case with sensor networks in real-world applications like crowdsourced earthquake detection or pollution monitoring. In response, we developed a new method that can effectively learn spatiotemporal dynamics from such point process observations. Our model integrates techniques from neural differential equations, neural point processes, implicit neural representations and amortized variational inference to model both the dynamics of the system and the probabilistic locations and timings of observations. It outperforms existing methods on challenging spatiotemporal datasets by offering substantial improvements in predictive accuracy and computational efficiency, making it a useful tool for modeling and understanding complex dynamical systems observed under realistic, unconstrained conditions.

Introduction: The Challenge of Random Observations

Predicting the evolution of spatiotemporal systems—such as weather patterns, wildfire spread, or fluid dynamics—is a cornerstone of modern scientific computing. Most existing deep learning approaches for these tasks rely on the assumption of "gridded" data: observations collected at fixed spatial locations and regular time intervals.

In practice, however, data is often "messy." Sensor networks, such as mobile phones or drifting ocean buoys, provide observations at random times and locations, forming what is known as a spatiotemporal point process. Existing models struggle with such sparse and irregularly sampled data, often failing to capture the underlying physics or the stochastic nature of the observation process itself. This paper introduces a framework designed to bridge this gap by jointly modeling the system's physics and the random process of observing it.

Methodology: A Generative Framework for Dynamics

We propose a generative framework that treats the system as a marked spatiotemporal point process. Our architecture is built upon four foundational pillars:

  1. Latent Dynamics via Neural ODEs: To handle the continuous nature of time, we compress the system state into a low-dimensional latent vector $z(t)$. Its evolution is governed by an Ordinary Differential Equation (ODE) parameterized by a neural network, allowing for continuous-time modeling.
  2. Implicit Neural Representations (INR): We decouple the latent state from any specific spatial grid. A Multi-Layer Perceptron (MLP) acts as an INR, mapping the latent vector and a spatial coordinate $x$ to the physical state $u(t, x)$. This makes the model resolution-independent.
  3. Neural Point Processes: We model the observation process using a non-homogeneous Poisson process. The intensity of observations is learned as a function of the latent state, enabling the model to understand if certain physical conditions (e.g., high temperature) correlate with more frequent measurements.
  4. Efficient Interpolation: A key technical innovation is the use of an auxiliary grid for ODE solving. Instead of solving the ODE at every unique observation timestamp—a major computational bottleneck—we solve it on a sparse grid and use linear interpolation to find states at specific observation times.
The proposed architecture integrates an encoder for context processing, Neural ODEs for continuous latent dynamics, and Implicit Neural Representations for spatial mapping.

Experimental Evaluation

The model was evaluated against several baselines, including Graph Neural Networks (FEN) and specialized point process models (DSTPP, AutoSTPP), across four diverse datasets: Burgers’ Equation (1D waves), Shallow Water Equations (2D waves), Navier-Stokes (2D fluid flow), and Scalar Flow (real-world 3D smoke plumes).

Key Results

Our framework consistently achieved lower Mean Absolute Error (MAE) and higher log-likelihoods than competing methods. Notably, the "interpolation trick" allowed for a 9x speedup in latent state evaluation and a 4x reduction in total training time without sacrificing accuracy. Furthermore, we found that low-dimensional latent spaces (16–64 dimensions) were sufficient to capture the dynamics of complex systems, suggesting that the underlying physics resides on a relatively simple manifold.

Qualitative results demonstrate the model's ability to reconstruct continuous physical fields from highly sparse and random observations (indicated by 'x' markers).

Discussion and Conclusion

By jointly modeling the system's physics and the random process of observation, our approach provides a robust solution for real-world spatiotemporal modeling. The combination of Neural ODEs and INRs offers a flexible, resolution-independent framework that is both accurate and computationally efficient. Future research could extend this work to scenarios where the act of observation influences the system's dynamics or where sensors actively interact with their environment.

Summary updated

Abstract

Spatiotemporal dynamics models are fundamental for various domains, from heat propagation in materials to oceanic and atmospheric flows. However, currently available neural network-based spatiotemporal modeling approaches fall short when faced with data that is collected randomly over time and space, as is often the case with sensor networks in real-world applications like crowdsourced earthquake detection or pollution monitoring. In response, we developed a new method that can effectively learn spatiotemporal dynamics from such point process observations. Our model integrates techniques from neural differential equations, neural point processes, implicit neural representations and amortized variational inference to model both the dynamics of the system and the probabilistic locations and timings of observations. It outperforms existing methods on challenging spatiotemporal datasets by offering substantial improvements in predictive accuracy and computational efficiency, making it a useful tool for modeling and understanding complex dynamical systems observed under realistic, unconstrained conditions.
BibTeX
@inproceedings{iakovlev2025modeling,
  author = {Valerii Iakovlev and Harri Lähdesmäki},
  title = {{Modeling Randomly Observed Spatiotemporal Dynamical Systems}},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/forum?id=37EXtKCOkn},
}
Figure