Machine Learning for fluid velocimetry and visualization
In the recent past, serval Convolutional Neural Networks (CNNs) have been developed to process Particle Image Velocimetry (PIV) and Background Oriented Schlieren (BOS) recordings. Despite showing promising results, several bottlenecks prevent their spread in the community. For instance, obtaining training data from Direct Numerical Solutions (DNS) for many flow regimes and geometry can be prohibitive, in terms of computational cost. Moreover, the architectures that have been proposed are very large, and thus require powerful GPU hardware and long computational times to process the recordings.
At our lab, we have developed an innovative training strategy which relies on kinematic motion. More precisely, we generate training sets in which particles move according to a random displacement field. This allows us to generate training data with a large variation of spatial scales at little computational cost, thus allowing for a higher generalization of the CNN.
Secondly, we also developed a novel CNN architecture, namely Lightweight Image Matching Architecture (LIMA), that has a considerably lower number of parameters compared to existing networks. LIMA can process images faster than other ML and classical approaches at comparable or higher accuracy. The network size is small enough that it can be compiled and deployed to embedded GPU devices, paving the way to the development of autonomous measurement systems.
