Dual-directional small-sampling deep-learning modelling on co-flowing microfluidic droplet generation

Abstract

User-specified droplets generated by microfluidics are critical but requires intensive expertise and much time. Stimulated by data boosts from microfluidic experiments, deep learning has proven to be a powerful modeling approach in microfluidics with high accuracy. However, current deep learning approaches on microfluidics, emphasize droplet size prediction rather than actually obtaining droplets of the user-desired size. Such droplet size predictions invariably overlook the effect of the flow regime, a critical factor when determining the availability of the generated droplets because the flow regime influences the droplets’ dispersity. In addition, the big data prerequisite in deep learning models prohibits its extensive application in microfluidics. In order to solve these problems, this paper combines our current extensive experimental data (extracted by the Hough transformer) and the experimental data from previous literature as a co-flowing microfluidics-based droplet database. Based on this database, we established a dual-directional deep learning model where the droplet size variable resides not only in the output (for prediction), but also in the input (for acquisition) when considering the effect of flow regime. The new integrated active learning model fundamentally reduces the training dataset without sacrificing accuracy, thus pioneering small-sampling deep learning modelling in microfluidics. Compared to the single-directional modelling approach, the proposed dual-directional modelling demonstrates obvious improvements where the average relative error is only 9.90 % (obtained droplet size compared to the desired size), an increase of 45.6% compared with that of the single-directional model. The improved deep learning methodology here offers a universally accurate model for co-flowing microfluidic-based droplet acquisitions. It also has great prospects of becoming a prominent data processing structure that boosts, and possibly transforms existing microfluidic research and related fields.