Image-Based Soft-Sensor Approach to Automate Microfluidic Control

Leveraging open-source AI platforms, this technology enables automated image-based detection, prediction, and control of microfluidic single bubble/droplet formation.
Problem:
Given the incorporation of microfluidic bubble/droplet generation in many applications of science and medicine, improving their high-quality throughout with minimal human intervention is crucial. There is currently a lack of tools available to automate microfluidic control, which is critical to decrease dependency on human expertise and scale the overall throughput and implementation of microfluidic devices.
Solution:
To maximize the capabilities of microfluidic technologies, the inventors develop a two-step soft-sensor (virtual-sensor) and respond strategy to enable an autonomous microfluidic system using a combination of image-based AI and air pressure control to correct droplet formation.
Technology:
To decrease the levels of necessary human intervention, this technology leverages open-source AI softwares to develop a convolutional neural network (CNN) trained on thousands of images to predict and correct aberrant droplet formations. This is achieved using a two-step soft-sensor approach that queries the droplet formation using a high-speed camera and implements an image recognition algorithm and CNN for feature extraction. These inform the proportional-integral-derivative (PID) controller to establish set-point tracking and rejection with direct air pressure modulation at the site of droplet formation on the device to achieve long-term droplet stability with minimal human intervention.
Advantages:

CNN was trained on 40,000+ 128 x 600 resolution images and validated on 4,800+ images.
CNN accuracy tests reveal 100% accuracy out of 450 images of novel test data (outside of the validation set).
Image algorithm incorporates an assessment of the (1) flow rate, (2) size, and (3) uniformity of the bubbles/droplets formed.
PID Feedback Control (in response to CNN detection) is enabled using either aqueous flow rate or gas pressure to maintain the desired bubble/droplet size setpoint.
This system demonstrates robust disturbance rejection, enabling over 99.2% (compared to 2.16% without the controller) of bubbles produced being within 5% of the setpoint value over an 8-hour production time.

Stage of Development:

Proof of Concept
Bench Prototype

Control schematic using CNN & image recognition for PID control of microfluidic bubble generation. The high-speed camera senses and relays the droplet size and shape, which is then fed through the CNN and overlayed onto the desired set point. The discrepancy is then used to compute the amount of pressure needed to be increased or decreased through the PID in order to maintain the set-point.
Intellectual Property: