Droplet menisci recognition by deep learning for digital microfluidics applications

Abstract

This paper demonstrates the use of deep learning, specifically the U-Net model, to recognize the menisci of droplets in an electrowetting-on-dielectric (EWOD) digital microfluidic (DMF) device. Accurate recognition of droplet menisci would enable precise control over the movement of droplets to improve the performance and reliability of an EWOD DMF system. Furthermore, important information such as fluid properties, droplet characteristics, spatial position, dynamic behavior, and reaction kinetics of droplets during DMF manipulation can be understood by recognizing the menisci. Through a convolutional neural network utilizing the U-Net architecture, precise identification of droplet menisci is achieved. A diverse dataset is prepared and used to train and test the model. As a showcase, details of training and the optimization of hyperparameters are described. Experimental validation demonstrated that the trained model achieves a 98% accuracy rate and a 0.92 Dice score, which confirms the model's high performance. After the successful recognition of droplet menisci, post-processing techniques are applied to extract essential information such as the droplet and bubble size and volume. This study shows that the trained U-Net model is capable of discerning droplet menisci even in the presence of background image interference and low-quality images. The model can detect not only simple droplets, but also compound droplets of two immiscible liquids, droplets containing gas bubbles, and multiple droplets of varying sizes. Finally, the model is shown to detect satellite droplets as small as 2% of the size of the primary droplet, which are byproducts of droplet splitting.