TLDR: DeepUFNet, a new AI model combining UNet and Fourier neural networks with a physics-aware loss control, accurately forecasts spatiotemporal wall pressure on rectangular cylinders. Trained on wind tunnel data, it shows high agreement with experimental results in various aspects and demonstrates strong extrapolation capabilities even with sparse input data, offering a robust solution for engineering applications.
Understanding how fluids exert pressure on structures is crucial for designing safe buildings and other engineering marvels. Traditional methods and even earlier deep learning approaches often struggle with predicting the complex, ever-changing (spatiotemporal) pressure across a surface, especially when dealing with limited data.
A new study introduces DeepUFNet, a cutting-edge deep learning model designed to accurately forecast spatiotemporal wall pressure on structures like rectangular cylinders. This innovative model combines the strengths of two powerful neural network architectures: the UNet structure and the Fourier neural network. The UNet is excellent at capturing hierarchical features, while the Fourier neural network excels in processing frequency-domain information, which is vital for understanding fluctuating pressures.
What makes DeepUFNet particularly advanced is its “physics-aware” design. During the model’s training, a unique high-frequency loss control mechanism, governed by a dynamic coefficient called beta, is embedded. This coefficient adjusts its influence as training progresses, guiding the model to better understand and predict high-frequency fluctuations in pressure, which are often critical for structural response.
To train and test DeepUFNet, researchers conducted detailed wind tunnel experiments. They collected high-frequency wall pressure data from a two-dimensional rectangular cylinder, creating a robust dataset. The model was then evaluated against this experimental data across various metrics.
The results were highly promising. DeepUFNet demonstrated remarkable accuracy in forecasting spatiotemporal wall pressure. Comparisons showed strong agreement between the model’s predictions and experimental data in terms of statistical information, how pressure changes over time, the power spectrum density (which reveals dominant frequencies), spatial distribution, and even spatiotemporal correlations. The inclusion of the dynamic beta coefficient was found to significantly enhance the model’s performance, particularly in forecasting high-order frequency fluctuations and pressure variance.
Beyond its forecasting accuracy, DeepUFNet also proved its ability to extrapolate. Even when provided with sparse spatial information (fewer measurement points), the model could still satisfactorily forecast the full spatiotemporal wall pressure. This capability is incredibly valuable for real-world engineering scenarios where complete sensor data might be unavailable or corrupted.
Also Read:
- Unlocking Efficient 3D Electromagnetic Scattering Modeling
- Accelerating PDE Simulations with Trainable-by-Parts AI
This research marks a significant step forward in applying artificial intelligence to fluid dynamics, offering a robust tool for predicting complex pressure phenomena. For more in-depth information, you can refer to the full research paper available here.
