Artificial Intelligence-Empowered Automated Double Emulsion Droplet Library Generation

Abstract

Double emulsions with core-shell structures are versatile materials used in applications such as cell culture, drug delivery, and materials synthesis. A droplet library with precisely controlled dimensions and properties would streamline screening and optimization for specific applications. While microfluidic droplet generation offers high precision, it is typically labor-intensive and sensitive to disturbances, requiring continuous operator intervention. To address these limitations, we present an artificial intelligence (AI)-empowered automated double emulsion droplet library generator. This system integrates a convolutional neural network (CNN)-based object detection model, decision-making, and feedback control algorithms to automate droplet generation and collection. The system monitors droplet generation every 171 ms-faster than a Formula 1 driver's reaction time-ensuring rapid response to disturbances and consistent production of single-core double emulsions. It autonomously generates libraries of 25 distinct monodisperse droplets with user-defined properties. This automation reduces labor and waste, enhances precision, and supports rapid and reliable droplet library generation. We anticipate that this platform will accelerate discovery and optimization in biomedical, biological, and materials research.

Keywords: convolutional neural network; experiment automation; feedback control; microfluidics; object detection.

Figure 1

Automated droplet library generator (ADLib). a) Schematic illustration of ADLib, consisting of: (1 and 2) a graphical user interface (GUI) on a computer; (3–5) three syringe pumps for the inner phase 1 and 2, and the middle phase; (6) a pressure controller and reservoir for the outer phase; (7) a high‐speed camera installed on an optical microscope; (8) a microfluidic device incorporating a droplet generator and a microfluidic mixer chip; (9) a product selector; and (10) a rotational collector. b) Operation of the automated droplet library generator program: (i and ii) continuous real‐time monitoring of droplet generation images, and (iii–v) the three primary functions for the generation of a droplet library: (iii) regulation of droplet size by feedback control of the outer phase pressure, (iv) selective collection of single‐core double emulsion droplets by a decision‐making algorithm, and (v) adjustment of flow rates and vessel switching for generation and collection of distinct double emulsion droplets.

Figure 2

Fine‐tuning and performance evaluation of YOLOv10n object detection model. a) Classes of objects that are detected in different modes of double emulsion generation, along with examples of droplet generation images with annotations and illustrations of droplets produced in each generation mode. IP, MP, and OP represent the inner, middle, and outer phase fluids, respectively. The width of the droplet generation images is equivalent to 200 µm. b) Normalized confusion matrix of the fine‐tuned object detection model tested on the validation data. c) Comparison of the outer diameter (D_O ) and coefficient of variation (CV) of double emulsion droplets: i) approximated from the object detection (OD) results, and ii) determined using the circle fitting method (CFT). The width of the droplet generation images is equivalent to 256 µm. d) Histograms show the distribution of processing times for size calculation from detection results (t_cal ), OD (t_OD ), and CFT (t_CFT ), and the overall image processing time (t_tot ) for a single loop of droplet generation monitoring process. The numbers on the histogram are the mean value and standard deviation of the image processing times.

Figure 3

Single‐core double emulsion (SCDE) generation recovery process. a) Droplet generation images and detection results during the SCDE recovery process. Information on the top left and bottom left of each image displays the current step, time, and droplet generation mode, respectively. Flow rates for inner phase 1 and 2, and the middle phase are 1.7, 0.3, and 0.5 mL h⁻¹, respectively. Droplet generation undergoes the following changes due to unknown disturbance and a recovery process that is executed by the droplet library generation program: i) the generation and collection of SCDE droplets, ii) the deviation of the inner phase from the middle phase and detection of MP mode (t = t₀ ), iii) the restoration of the coaxial jet with the middle phase enveloping the inner phase by pulsing the middle phase flow, iv) the stabilization of the droplet generation system, leading to the generation of SCDE droplets, v) feedback control of D_O by tuning P_OP , vi) waiting for defects to flow out of the device, and vii) the resumption of SCDE droplets collection. The width of the droplet generation images is equivalent to 256 µm. b–d) Temporal changes in (b) the outer diameter of double emulsion droplet, (c) the detected droplet generation mode, (d) the outer phase pressure (P_OP ), and the middle phase flow rate (Q_MP ) during the SCDE recovery process.

Figure 4

Feedback control of double emulsion droplet size by adjusting the outer phase pressure. a) Flowchart of the droplet size feedback loop. Arrows indicate flow directions. The program compares the current outer diameter (D_O ) of the droplet with the target diameter (target D_O ) and updates the outer phase pressure (P_OP ) via a proportional controller with the proportional gain, K_P . The feedback loop continues updating P_OP until the outer diameter error (D_err ) is <2% of the target D_O . b) Temporal variation of P_OP adjusted by the feedback control, with a dispersed flowrate (Q_DP = Q_IP + Q_MP ) of 2.5 mL h⁻¹, and a target D_O of 55 µm. c–e) Inner and outer diameters (D_I and D_O ) and shell thickness (t_shell ) of double emulsion droplets as a function of the flow rate ratio of the middle phase to the inner phase (Q_MP /Q_IP ) with (c) and without (d, e) feedback control. (i, ii) Optical micrographs of the double emulsion droplets produced under (i) Q_MP /Q_IP = 0.25 and (ii) Q_MP /Q_IP = 1. The scale bar is 100 µm.

Figure 5

Generation of a double emulsion droplet library with varying solute concentration and shell thickness. a) Flow rates for inner phase 1 and 2, and the middle phase, used to generate a 5 × 5 double emulsion droplet library. The flow rates are calculated by the droplet library generation program based on the following inputs: target droplet size = 53 µm, shell thicknesses = [5, 4, 3, 2, 1.5] µm, solute concentration of inner phase 2 = 5 mg mL⁻¹, minimum and maximum solute concentration in the inner phase = [0.825, 4.125] mg mL⁻¹, and dispersed phase flowrate = 2.5 mL h⁻¹. b) Temporal variation of the outer phase pressure as adjusted by feedback control. c) Outer diameter (D_O ) and coefficient of variation (CV) of double emulsion droplet as functions of time. d) Changes in normalized mean gray value of the double emulsion droplet in high‐speed images over time. The term “Collect” and the subscript “Clt” represent data from images when droplets are collected, while “Waste” and the subscript “Wst” refer to data from images of droplets flowing into the waste vessel.

Figure 6

Double emulsion droplet library with varying concentration and shell thickness. a) Optical micrographs of double emulsion droplets from a 5 × 5 droplet library varying in the dye concentration and the shell thickness. The scale bar represents 100 µm. b) Outer diameter and c) shell thickness of the double emulsion droplets in the droplet library as a function of the flow rate ratio of the middle phase and the inner phase. “C” from the legend indicates the solute concentration in the inner phase of the double emulsion droplet.