Abstract
Reconstructing spatiotemporal dynamics with sparse sensor measurement is an outstanding problem, commonly encountered in a wide spectrum of scientific and engineering applications. Such a problem is particularly challenging when the number and/or types of sensors (e.g., randomly placed) are extremely insufficient. Existing end-to-end learning models ordinarily suffer from the generalization issue for full-field reconstruction of spatiotemporal dynamics, especially in sparse data regimes typically seen in real-world applications. To this end, we propose a sparse-sensor-assisted score-based generative model ($S^3GM$) to reconstruct and predict full-field spatiotemporal dynamics based on sparse measurements. Instead of learning directly the mapping between input and output pairs, an unconditioned generative model is firstly pretrained capturing the joint distribution of a vast group of pretraining data in a self-supervised manner, followed then by a sampling process conditioned on unseen sparse measurement. The efficacy of $S^3GM$ has been verified on multiple dynamical systems with various synthetic, real-world, and lab-test datasets (ranging from turbulent flow modeling to weather/climate forecasting). The results demonstrate the excellent performance of $S^3GM$ in zero-shot reconstruction and prediction of spatiotemporal dynamics even with high levels of data sparsity and noise. We find that $S^3GM$ exhibits high accuracy, generalizability, and robustness when handling different reconstruction tasks.
Introduction
In scientific and engineering disciplines such as meteorology, fluid mechanics, and aerospace engineering, the ability to reconstruct full-field dynamics from limited observations is crucial. However, real-world scenarios often provide only sparse, noisy, and irregularly placed sensor data. Traditional deep learning approaches typically rely on end-to-end mappings that are highly sensitive to sensor configurations, often failing when faced with unseen sensor types or placements. This lack of flexibility hinders their application in heterogeneous real-world environments.
The $S^3GM$ Framework
We present the Sparse-Sensor-Assisted Score-Based Generative Model ($S^3GM$), a framework designed to overcome the limitations of direct mapping by learning the underlying probability distribution of physical systems. Unlike models that bake physical equations into loss functions, $S^3GM$ learns implicit physics from high-fidelity data, providing greater robustness to varying system parameters.
The $S^3GM$ methodology involves two distinct stages:
- Self-Supervised Pretraining: A score-based generative model (SGM) is trained to learn the joint distribution of state variables and controlling parameters (e.g., viscosity). The architecture employs a "video U-Net" featuring spatial and temporal attention to capture complex physical patterns and temporal dependencies.
- Posterior Sampling for Reconstruction: For new tasks, the pretrained model performs zero-shot reconstruction by solving a stochastic differential equation (SDE). This process refines Gaussian noise into a physical field, constrained by observation consistency with sensor data and sequence consistency for temporal alignment.
Experimental Evaluation
The $S^3GM$ framework was rigorously tested across four regimes: chaotic systems (Kuramoto-Sivashinsky Equation), turbulent flows (Kolmogorov flow), real-world climate data (ERA5), and experimental lab data (cylinder flow).
Key Results
- Zero-Shot Generalization: The model successfully reconstructed fields using sensor types it never encountered during training, such as Fourier spectral data or time-averaged point measurements.
- Robustness to Noise: In turbulent flow scenarios, $S^3GM$ accurately recovered velocity fields from just 5% spatial sampling with 10% added Gaussian noise, significantly outperforming baselines like Fourier Neural Operators (FNO).
- Out-of-Distribution Performance: The model generalized to higher Reynolds numbers ($Re=1500$) than those seen during training ($Re \le 1050$), indicating it captures fundamental, scalable physical patterns.
- Climate Imputation: Using 1% noisy wind data, the model successfully imputed unobserved variables like temperature and surface pressure with high fidelity.
Conclusion
$S^3GM$ offers a flexible and robust solution for reconstructing spatiotemporal dynamics from sparse data. By leveraging generative priors, it achieves superior zero-shot generalization and noise resilience compared to traditional end-to-end models. While the computational cost of SDE sampling remains a challenge, the framework's ability to handle diverse and unseen sensor configurations marks a significant step forward in physical system modeling.
Summary updated