M.Sc. Kyriacos Chr. Yiannacou introduces a method for controlling the positions of particles and droplets inside microfluidic chips using ultrasound and machine learning algorithms. In his dissertation, Yiannacou presents a manipulation method for overcoming limitations of current acoustic manipulation methods.
Current methods often require precise modeling of the acoustic field shapes through calibration experiments or extensive simulations. These approaches do not always accurately represent reality and can lead to manipulation failures. Using machine learning algorithms – namely the ε-greedy, the upper confidence bound (UCB), and a new variation of the ε-greedy called the acoustic-model-based adaptive controller or AMA for short – Yiannacou showed real-time manipulations of a single, multiple particles and droplets without any prior calibration or simulations.
All algorithms were successful in completing different manipulation tasks. However, Yiannacou evaluated their performance in terms of their manipulation accuracy and speed.
“The AMA algorithm outperformed the UCB and ε-greedy algorithm in terms of manipulation speed and accuracy, especially when I repeated the task more than once. This was a result of the algorithms’ ability to learn the shape of the acoustic field for each frequency and adapt to use only the best ones for the manipulation task,” Yiannacou says.
Further, Yiannacou applied the algorithms we to various tasks that could represent real-world applications, such as sorting particles through different outlets and merging droplets inside the microfluidic chamber. This highlights that the microfluidic chip using these algorithms can be "re-programmed" and used for different applications each time.
Yiannacou’s dissertation bridges the fields of microfluidics, machine learning, and acoustics to create a microfluidic system with various applications. The adaptive machine learning algorithms help tackle disturbances that can change the shape of the acoustic fields. Examples of such disturbances include air bubbles entering the chamber or changes in fluid properties. However, the adaptiveness of the algorithms makes this approach highly robust.
“By integrating machine learning control methods with acoustic microfluidics, we can achieve online closed-loop control that can be used to flexibly manipulate particles and droplets inside microfluidic chips, making these systems more dependable and versatile,” Yiannacou says.
Public defence on Friday 7 February
The doctoral dissertation of M.Sc. Kyriacos Chrysanthos Yiannacou in the field of Biomedical Engineering titled Machine Learning-Based Acoustic Manipulation of Particles and Droplets inside Microfluidic Devices will be publicly examined at the Faculty of Medicine and Health Technology at Tampere University at 12 o’clock on Friday 7 February 2025 at Hervanta campus, Festia, Pieni Sali 1 (FA032) (Korkeakoulunkatu 8, Tampere). The Opponent will be Associate Professor Per Augustsson from Lund University, Sweden. The Custos will be Associate Professor Veikko Sariola from the Faculty of Medicine and Health Technology, Tampere University.