Shallow neural networks for fluid flow reconstruction with limited sensors

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

TL;DR

This paper introduces a "Shallow Decoder" (SD) neural network for reconstructing complex fluid flow fields from sparse sensor data. The SD learns an end-to-end mapping, outperforming traditional methods like Proper Orthogonal Decomposition (POD) in accuracy and requiring significantly fewer sensors, making it ideal for real-world monitoring applications.

What Problem Does This Paper Address?

Accurately understanding and predicting fluid flow is crucial for many applications, from designing more fuel-efficient vehicles to monitoring climate. However, obtaining high-resolution data of fluid flow fields is often expensive or impossible. This necessitates reconstructing the full flow field from a limited number of sensor measurements. Traditional methods, like POD, struggle with noisy or non-linear sensor data and often require a large number of sensors for good performance. This paper addresses the challenge of reconstructing complex, high-dimensional fluid flow fields accurately and efficiently using a minimal set of sensor measurements, even when the sensor data is non-linear or noisy.

Methods

The core contribution is a Shallow Decoder (SD), a type of neural network designed for this reconstruction task. Unlike traditional methods that first learn a basis (like POD modes) and then estimate coefficients, the SD learns an end-to-end mapping directly from sensor measurements to the reconstructed flow field. This approach is purely data-driven, requiring no prior knowledge of the underlying physics or measurement operator. The architecture is intentionally shallow, featuring two hidden layers and a linear output layer, which aids in faster training, easier interpretation, and better generalization. The network is trained to minimize the difference between the reconstructed and true flow fields. The authors also incorporate standard regularization techniques like batch normalization and dropout to improve robustness.

Datasets

The paper evaluates the SD on three distinct datasets:

  • Fluid flow behind a circular cylinder: A canonical example in fluid dynamics, used to test reconstruction with varying numbers of sensors, non-linear measurements, and noisy data.
  • Weekly mean sea surface temperature (SST): A more challenging, non-stationary dataset from NOAA, used to assess performance with random point-wise measurements.
  • Turbulent isotropic flow: A highly non-stationary and multiscale flow field, used to test reconstruction from sub-gridscale measurements.

Key Insights and Ideas

  • End-to-end learning simplifies reconstruction: By directly mapping sensor data to the flow field, the SD bypasses the multi-step process of traditional methods, leading to improved performance and efficiency.
  • Shallow networks are sufficient and interpretable: The authors argue that shallow architectures, with their inherent simplicity, are effective for this task and facilitate easier interpretation of the learned features.
  • Data-driven approach handles non-linearity and noise: The SD's ability to learn complex mappings makes it robust to non-linear sensor measurements and additive noise, a significant advantage over linear POD-based methods.
  • Sensor placement is less critical: While traditional methods are highly sensitive to sensor location, the SD demonstrates good performance even with random sensor distributions, reducing the burden of optimal sensor placement.
  • Extrapolation is a limitation: The SD, like most data-driven methods, performs well for interpolation (within the training data distribution) but struggles with extrapolation (predicting states far outside the training distribution), particularly for highly non-stationary flows.

Key Results

The SHALLOW DECODER consistently outperformed traditional POD-based methods across all tested scenarios:

  • Flow behind cylinder: Achieved excellent reconstruction with as few as 5 sensors, while POD methods required significantly more (e.g., >15) for comparable accuracy. The SD also showed robustness to non-linear measurements and noise, achieving very low NME and NFE errors (e.g., ~0.006 NME with 5 sensors).
  • Sea surface temperature: The SD demonstrated excellent reconstruction quality for fluctuations around the mean, with errors as low as 12% (NFE). It also showed high invariance to sensor location.
  • Turbulent flow: The SD yielded excellent results for high-dimensional flow fields even with challenging sub-gridscale measurements, with interpolation errors around 9.3%.
  • Training time: Training the SD was minimal, often taking less than two minutes on a modern GPU for the tested examples.

Limitations and Future Work

  • Extrapolation limitations: The paper explicitly notes that the SD cannot be reliably used for "out of sample prediction tasks" if the fluid flow is highly non-stationary. Extrapolation to states far outside the training data distribution leads to reconstruction failure.
  • Data representativeness: The SD, like all data-driven methods, requires training data that is representative of the system's statistical distribution.
  • Future Work: The authors suggest leveraging underlying physical laws to further improve efficiency and integrating insights from physics into the shallow decoder network. They also propose exploring the use of shallow decoders for reduced-order modeling (ROMs) for forecasting predictions, especially in high-dimensional systems where POD might be less effective.

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 to 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 neural networks for fluid flow reconstruction with limited sensors}},
  booktitle = {Proceedings of the Royal Society A Mathematical Physical and Engineering Sciences},
  year = {2020},
  doi = {10.1098/rspa.2020.0097},
}
Figure