Reproduce “Enhancing wind field resolution in complex terrain through a knowledge-driven machine learning approach” on OpenResearch

About this paper

Jacob Wulff Wold, Florian Stadtmann, Adil Rasheed, Mandar Tabib, Omer San, Jan-Tore Horn

Researchers developed a knowledge-driven generative convolutional network to super-resolve 3D wind fields in complex terrain, achieving an absolute pixel error of 0.116 m/s compared to 0.385 m/s for trilinear interpolation. The approach incorporates physics-based loss functions, eliminating the need for adversarial training, and includes the release of a 2TB high-resolution wind field dataset.

Key insights

• Physics-based loss components, specifically those derived from wind field gradients and divergence, are crucial for robust training and accuracy in super-resolving complex 3D wind fields.
• The integration of physics-informed loss functions can stabilize the training process and achieve high performance, effectively mitigating or eliminating the need for adversarial learning components in GANs.
• Including auxiliary input features such as height above mean sea level or terrain height consistently improves the accuracy of wind field super-resolution.

Summary by alphaXiv.