Efficient deep data assimilation with sparse observations and time-varying sensors

Sibo Cheng, Che Liu, Yike Guo, Rossella Arcucci

Abstract

Variational Data Assimilation (DA) has been broadly used in engineering problems for field reconstruction and prediction by performing a weighted combination of multiple sources of noisy data. In recent years, the integration of deep learning (DL) techniques in DA has shown promise in improving the efficiency and accuracy in high-dimensional dynamical systems. Nevertheless, existing deep DA approaches face difficulties in dealing with unstructured observation data, especially when the placement and number of sensors are dynamic over time. We introduce a novel variational DA scheme, named Voronoi-tessellation Inverse operator for VariatIonal Data assimilation (VIVID), that incorporates a DL inverse operator into the assimilation objective function. By leveraging the capabilities of the Voronoi-tessellation and convolutional neural networks, VIVID is adept at handling sparse, unstructured, and time-varying sensor data. Furthermore, the incorporation of the DL inverse operator establishes a direct link between observation and state space, leading to a reduction in the number of minimization steps required for DA. Additionally, VIVID can be seamlessly integrated with Proper Orthogonal Decomposition (POD) to develop an end-to-end reduced-order DA scheme, which can further expedite field reconstruction. Numerical experiments in a fluid dynamics system demonstrate that VIVID can significantly outperform existing DA and DL algorithms.

Introduction

Data Assimilation (DA) is a critical technique for estimating the true state of a system by combining numerical models with real-world observations. While traditional variational DA is physically consistent, it is computationally expensive, often requiring numerous iterations to solve complex optimization problems. Deep Learning (DL) has emerged as a potential solution to accelerate this process; however, standard DL architectures like Convolutional Neural Networks (CNNs) typically require structured sensor grids. This limitation makes them difficult to apply to real-world scenarios where sensors are sparse, unstructured, or mobile (e.g., ocean buoys).

The VIVID Framework

To address these challenges, we propose VIVID (Voronoi-tessellation Inverse operator for VariatIonal Data assimilation). The framework consists of three primary components designed to bridge the gap between unstructured data and structured DL models:

  1. Voronoi Tessellation: To handle sensors at arbitrary locations, the observation space is partitioned into cells based on proximity to each sensor. This creates a tessellated field that can be processed by standard CNNs.
  2. DL Inverse Operator: A CNN is trained to map these tessellated observations directly back to the full state space, providing a global prediction based on historical data.
  3. Hybrid Objective Function: VIVID integrates the DL inverse operator directly into the DA cost function. This allows the optimization algorithm to find solutions in significantly fewer steps by establishing a direct link between observation and state space.
Schematic of the VIVID framework: Voronoi tessellation of sparse observations followed by a CNN-based inverse operator to reconstruct the system state.

The specific architecture of the CNN used in VIVID involves multiple 2D convolutional layers with ReLU activations, as detailed in the table below.

Architecture of the CNN layers used in the VIVID inverse operator.

Experimental Evaluation

The framework was evaluated using a Shallow Water CFD model (Saint-Venant equations). The model was trained on 10,000 snapshots from simulations with varying physical parameters and tested on an unseen simulation to assess generalizability. Synthetic sensors were placed randomly at each time step to simulate sparse, time-varying data.

Key Results

  • Accuracy: VIVID reduced the Relative Root Mean Square Error (R-RMSE) by approximately 50% compared to both traditional variational DA and standalone Voronoi-CNN (VCNN) approaches.
  • Computational Efficiency: In full physical space, VIVID was roughly 4x faster than traditional DA.
  • Reduced-Order Speed: The integration of Proper Orthogonal Decomposition (VIVID-ROM) further accelerated the process, reducing computation time from 42 minutes (traditional DA) to just 2 seconds.
  • Robustness: VIVID demonstrated high visual fidelity in capturing complex wave patterns and remained robust even when error covariances were misspecified.

Conclusion

VIVID successfully fuses the strengths of traditional DA and deep learning, offering a sensor-agnostic approach that is both fast and accurate. Future work will focus on incorporating explicit physical constraints and extending the framework to handle temporal sequences using architectures like Transformers or ConvLSTMs.

Summary updated

Abstract

Variational Data Assimilation (DA) has been broadly used in engineering problems for field reconstruction and prediction by performing a weighted combination of multiple sources of noisy data. In recent years, the integration of deep learning (DL) techniques in DA has shown promise in improving the efficiency and accuracy in high-dimensional dynamical systems. Nevertheless, existing deep DA approaches face difficulties in dealing with unstructured observation data, especially when the placement and number of sensors are dynamic over time. We introduce a novel variational DA scheme, named Voronoi-tessellation Inverse operator for VariatIonal Data assimilation (VIVID), that incorporates a DL inverse operator into the assimilation objective function. By leveraging the capabilities of the Voronoi-tessellation and convolutional neural networks, VIVID is adept at handling sparse, unstructured, and time-varying sensor data. Furthermore, the incorporation of the DL inverse operator establishes a direct link between observation and state space, leading to a reduction in the number of minimization steps required for DA. Additionally, VIVID can be seamlessly integrated with Proper Orthogonal Decomposition (POD) to develop an end-to-end reduced-order DA scheme, which can further expedite field reconstruction. Numerical experiments in a fluid dynamics system demonstrate that VIVID can significantly outperform existing DA and DL algorithms. The robustness of VIVID is also accessed through the application of various levels of prior error, the utilization of varying numbers of sensors, and the misspecification of error covariance in DA.
BibTeX
@article{cheng2024efficient,
  author = {Sibo Cheng and Che Liu and Yike Guo and Rossella Arcucci},
  title = {{Efficient deep data assimilation with sparse observations and time-varying sensors}},
  journal = {Journal of Computational Physics},
  year = {2024},
  url = {https://openreview.net/forum?id=YUENLYhm4r},
}
Figure