Using deep learning to recognize liquid–liquid flow patterns in microchannels

In this work, an automatic liquid–liquidtwo-phase flow pattern recognition platform was developed to help circumvent the difficulties in labor-intensive hydrodynamics studies. Trained by about 30,000 of human-labeled flow pattern images, a convolutional neural network was built with the expert-level ability in the flow pattern recognition tasks and then coupled with automatic pump feeding system and online high-speed camera monitoring system to realize the high-throughput experimentation platform for microchannels. Effects of important factors such as flow rate, viscosity, interfacial tension, and so on were studied, and different flow pattern maps were obtained. With these thousands of flow pattern data in hand, we eventually drew the generalized liquid–liquidtwo-phase flow map and then proposed the relatively prudent criteria for slug flow operation window in the microchannel. This study extended the applications of artificial intelligence on microreactor technology or microfluidics, and in particular facilitated understanding complex hydrodynamics and flow patterns.