Machine Learning Based Acoustic Manipulation of Particles and Droplets inside Microfluidic Devices

Abstract

This thesis introduces a novel approach to ultrasonically manipulate particles and droplets inside microfluidic devices using machine learning based control algorithms and bulk acoustic waves. Most reported techniques utilize one or more piezoelectric transducers to drive microfluidic chip channels primarily at their first resonance frequency or at most few harmonics. This causes excitation of acoustic modes within the channels that can move particles toward areas of minimum or maximum pressure within the channels. However, previously reported methods depend heavily on the careful design of the microfluidic chips and precise modelling of acoustic fields from calibration experiments or simulations; a task that becomes exceedingly complex with complex channel structures or when unexpected disturbances interfere with the modelled acoustic fields. Furthermore, when experimentally modelling the acoustic fields with particle tracking methods poses practical difficulties within a microfluidic chip, due to issues arising from particle clustering, and due to particles moving in and out of focus while imaged with a microscope.

This thesis investigates the application of machine learning algorithms to address these challenges when performing a two-dimensional (2D) manipulation of particles in microfluidic chips. Two microfluidic chips were used for this thesis, one with one outlet and another with three outlets, both chips included a rectangular shape chamber (manipulation chamber), where the 2D manipulation takes place. Three algorithms—the ε-greedy, UCB, and a new variant of the ε-greedy which we developed and call the acoustic- model-based adaptive controller, or AMA controller for short—were studied for their ability to manipulate droplets and particles within a microfluidic chip. The three algorithms learn how to manipulate the particles in an online manner. The algorithms were capable to adapting to sudden changes in the system, e.g., when an air bubble entered the manipulation chamber. Using these manipulation methods, we reported successful manipulation of a single and multiple particles, manipulation of immiscible droplets, particle sorting, and performing the chemistry underlying a colorimetric glucose assay by merging two droplets inside the chamber. From the three algorithms, the new AMA controller was found to perform the best in terms of manipulation speed and accuracy, especially in subsequent cycles of the same task when the controller had learned the acoustic field shapes, whereas the two other algorithms had comparable performance between them, but worse than the AMA algorithm overall.

To conclude, this thesis shows that machine learning algorithms are suitable for manipulating particles and droplets within microfluidic chips without calibration experiments, or simulations to obtain the acoustic field shapes, and can complete a manipulation task even on the first try. The proposed algorithms can adapt to dynamic changes, in subsequent cycles the developed AMA control algorithm improves its manipulation accuracy and speed. This approach signifies a promising direction for more precise and flexible control of particles and droplets in complex microfluidic systems.