From: 2025-08-15 10:30 to 11:30
Place: Seminar Room M 3170-73 in the M-building, LTH
Contact: anders [dot] rantzer [at] control [dot] lth [dot] se

Date & Time: August 15th, 10:30-11:30
Location: Seminar Room M 3170-73 in the M-building, LTH
Author: Guilherme J. Perticarari
Title: Acoustophoresis-driven Multiparticle Control in a Microfluidic Device
Supervisor:  Thierry Baasch (supervisor), Dongjun Wu (cosupervisor)
Examiner:  Anders Rantzer

Abstract: When particles are dispersed in a microfluidic device, its resonant frequencies can be used to manipulate their positions at will. This practice, called acoustophoresis, has been shown to successfully control multiple particles simultaneously with a single vibrating actuator, which is a low-cost and contactless alternative to existing particle manipulation methods, and whose applications include isolating Circulating Tumor Cells from blood samples. This work is an investigation into when and how this system can be controlled, with the former leading to a theoretical investigation on the system's controllability under varying numbers of particles and actuation frequencies, and the latter, to novel local optimization algorithms that achieve state-of-the-art solutions in simulated environments.
Our theoretical results show that the success rate of a random manipulation task is related to the ratio of the number of particles and frequencies, which means that a system with double the amount of particles, P, will maintain its success rate, R, if the amount of frequencies, M, is also doubled. In 1D problems with ordered frequencies, we show that the simple heuristic R(M,P) = 1-P/M predicts the system's controllability level reasonably well and, for more realistic systems, the controllability can be approximated using Wendel's theorem. Moreover, we find that M >4P seems to be a requirement for a controllability level above 50%. We also introduce three model-based algorithms tha outperform state-of-the-art alternatives in our numerical simulations. These algorithms were able to reach approximately ~80% of success rate in tasks involving up to four particles, while state-of-the-art solutions like Ɛ-greedy achieved ~60%. These novel models are able to 'learn-on-the-go', and, since they are model-based, they can be used to steer the particles towards different directions on a whim, without any need for retraining.