Thermal conductivity properties of certain colloidal suspensions, or nanofluids, make them useful substances for a variety of heat exchange processes. Using them effectively, however, demands an understanding of flow characteristics. Flow pattern recognition in gas-liquid structures, for instance, can be used in predicting heat transfer rates and equipment performance.
Identifying which of the numerous possible flow patterns such a two-phase fluid exhibits in a microchannel is sufficiently complex and difficult that the best option is typically direct, visual assessment. But with modern computing power, artificial intelligence continues to find more applications — including flow pattern identification. An artificial intelligence system described in AIP Advances identifies nanofluid gas-liquid two-phase flow with accuracy improved by high-speed imaging that feeds into its machine learning algorithms.
Images of nanofluids contained in vertical minichannels approximately 2 millimeters wide capture patterns of flow boiling with four different types of flow: annular flow, bubbly flow, churn flow and slug flow. Flow pattern recognition is realized using support vector machine (SVM) classification, in which supervised learning models analyze images captured with a high-speed camera. Textures in multi-frame sets of these images identify each flow-specific pattern. To improve recognition speed, real-time detection and principle component analysis (PCA) reduce dimensionality.
The system, which combines PCA with SVM, is not only faster, but more accurate than similar image based approaches using only SVM. For each of the four flow types, authors report over 90 percent accuracy. This approach offers improvements in controlling dynamics such as bubble activity which are important, for example, to hydrogen production technologies and droplet-based microfluidics.
Source: “Using artificial intelligence to improve identification of nanofluid gas–liquid two-phase flow pattern in mini-channel,” by Jian Xiao, Xiaoping Luo, Zhenfei Feng, and Jinxin Zhang, AIP Advances (2017). The article can be accessed at https://doi.org/10.1063/1.5008907.
