Learning to Reconstruct Flow Fields from Limited Measurements

We study the problem of perception and reconstruction of flow fluids with limited measurements. Deep learning-based approaches have outperformed traditional methods in this setting due to the capacity and generalization of deep learning-based approaches. It still lacks a principle for designing neural architectures. Inspired by Erichson [1], this paper takes multilayer perceptron as an example and explores the neural architecture search under the guidance of traditional principles. Lots of experiments have been conducted and the results demonstrate the validity of the neural architecture search.