Shallow Learning for Fluid Flow Reconstruction with Limited Sensors and Limited Data

N. Benjamin Erichson, Lionel Mathelin, Zhewei Yao, Steven L. Brunton, Michael W. Mahoney, J. Nathan Kutz

Abstract

In many applications, it is important to reconstruct a fluid flow field, or some other high-dimensional state, from limited measurements and limited data. In this work, we propose a shallow neural network-based learning methodology for such fluid flow reconstruction. Our approach learns an end-to-end mapping between the sensor measurements and the high-dimensional fluid flow field, without any heavy preprocessing on the raw data. No prior knowledge is assumed to be available, and the estimation method is purely data-driven. We demonstrate the performance on three examples in fluid mechanics and oceanography, showing that this modern data-driven approach outperforms traditional modal approximation techniques which are commonly used for flow reconstruction. Not only does the proposed method show superior performance characteristics, it can also produce a comparable level of performance with traditional methods in the area, using significantly fewer sensors. Thus, the mathematical architecture is ideal for emerging global monitoring technologies where measurement data are often limited.

The Challenge of Fluid Flow Reconstruction

In physical and engineering sciences—ranging from automotive design to oceanography—understanding the full state of a fluid flow is essential. However, practical constraints often limit researchers to a handful of sparse sensor measurements.

The traditional approach to this problem utilizes modal approximation techniques, most notably Proper Orthogonal Decomposition (POD). POD identifies a low-dimensional basis of dominant patterns from historical data and then uses least-squares regression to fit live sensor readings to these patterns. A critical weakness of this method is observability: the basis patterns are learned independently of sensor placement. If sensors are located in "blind spots" relative to these patterns, reconstruction fails. Additionally, traditional methods struggle with noisy or non-linear sensor data.

The Shallow Decoder Architecture

To overcome these limitations, we introduce the Shallow Decoder (SD), a purely data-driven framework that learns an end-to-end mapping from sparse sensor inputs directly to the high-dimensional flow field.

Overview of the Shallow Decoder (SD) framework for reconstructing high-dimensional flow fields from sparse sensor measurements.

The SD architecture is intentionally streamlined:

  1. Input Layer: Receives sparse sensor measurements.
  2. Hidden Layers: Two fully connected layers with ReLU activations extract non-linear features.
  3. Output Layer: A linear layer reconstructs the full high-dimensional field.

By using fully connected layers rather than convolutions, the SD remains agnostic to the spatial ordering of sensors and trains rapidly. Regularization techniques like Batch Normalization and Dropout are employed to prevent overfitting on small datasets.

Experimental Evaluation

The SD was evaluated across three datasets of increasing complexity:

  • Cylinder Flow: A 2D simulation of periodic vortex shedding behind a cylinder.
  • Sea Surface Temperature (SST): 26 years of global NOAA data featuring complex seasonal and non-stationary dynamics.
  • Isotropic Turbulence: A 3D chaotic flow from the Johns Hopkins Turbulence Database.

Key Insights and Results

The primary advantage of the SD is joint learning. By training end-to-end, the network simultaneously learns the approximation basis and the sensor mapping. This ensures the learned patterns are inherently observable by the specific sensor configuration.

Comparison of (a) Proper Orthogonal Decomposition (POD) modes and (b) modes learned by the Shallow Decoder output layer.

Key performance highlights include:

  • Efficiency: On cylinder flow, the SD achieved high accuracy with only 5 sensors, whereas POD required 15 and showed high sensitivity to placement.
  • Robustness: The SD acted as an effective denoiser. With a Signal-to-Noise Ratio of 10, POD failed while the SD maintained a relative error of ~27%.
  • Real-World Performance: On the SST dataset, the SD achieved a 12% error rate using 64 random sensors, significantly outperforming regularized POD.
  • Interpretable Physics: Despite being data-driven, the modes learned in the SD's final layer (Figure 5b) represent physically consistent flow structures, similar to traditional modes but optimized for the sensor layout.

Conclusion and Future Directions

The Shallow Decoder demonstrates that shallow neural networks are sufficient for high-fidelity flow reconstruction, offering a simpler and more robust alternative to deep convolutional models or traditional linear methods. While the current model is purely data-driven and may struggle with out-of-sample extrapolation, future work involves integrating physical laws (e.g., Navier-Stokes equations) into the loss function. Such physics-informed architectures could further enhance generalization and enable the development of predictive Reduced Order Models (ROMs).

Summary updated

Abstract

In many applications, it is important to reconstruct a fluid flow field, or some other high-dimensional state, from limited measurements and limited data. In this work, we propose a shallow neural network-based learning methodology for such fluid flow reconstruction. Our approach learns an end-to-end mapping between the sensor measurements and the high-dimensional fluid flow field, without any heavy preprocessing on the raw data. No prior knowledge is assumed to be available, and the estimation method is purely data-driven. We demonstrate the performance on three examples in fluid mechanics and oceanography, showing that this modern data-driven approach outperforms traditional modal approximation techniques which are commonly used for flow reconstruction. Not only does the proposed method show superior performance characteristics, it can also produce a comparable level of performance with traditional methods in the area, using significantly fewer sensors. Thus, the mathematical architecture is ideal for emerging global monitoring technologies where measurement data are often limited.
BibTeX
@inproceedings{erichson2020shallow,
  author = {N. Benjamin Erichson and Lionel Mathelin and Zhewei Yao and Steven L. Brunton and Michael W. Mahoney and J. Nathan Kutz},
  title = {{Shallow Learning for Fluid Flow Reconstruction with Limited Sensors and Limited Data}},
  booktitle = {Proceedings of the Royal Society A Mathematical Physical and Engineering Sciences},
  year = {2020},
  doi = {10.1098/rspa.2020.0097},
}
Figure