Acoustic Particle Control with Particle-Particle Interactions

Abstract

Acoustophoresis, the manipulation of microscopic particles by acoustic forces, has found growing use in biomedical applications ranging from blood-cell separation to single-cell positioning in microfluidic chips [1,2,3,4]. Existing multimodal control algorithms for such devices were developed under the assumption that particles do not influence one another. This simplification holds when only one particle is present or when particles are widely separated. In scenarios where particle trajectories can cross each other, or particles are present in close proximity, this simplification breaks down.

This thesis extends and evaluates a suite of acoustophoretic control algorithms, originally designed for non-interacting particles by Perticarari [5] and Öhrnberg [6], in the presence of acoustic and hydrodynamic particle-particle interactions. The work has two parts. In the first, the algorithms are re-benchmarked in the non-interaction setting. In the second, the same algorithms are evaluated inside an interaction-enabled simulation framework, where the Acoustic Interaction Force (AIF), hydrodynamic coupling between particles, and contact forces are fully accounted for.

Five controllers are studied: the epsilon-greedy algorithm [7], the model-based quadratic programme, here named (QP) or as named by Perticarari [5] Perfect-Information Local Optimisation (PILOt), the online-learning Vector-based Local Optimisation (VeLO) controller and its no-sweep variant, and a tile-coded tabular Q-learning. To ensure consistent and reproducible evaluation, a deterministic main benchmark consisting of a varied number of samples under multiple environment setups is developed and applied across all controllers.

The results show that QP, VeLO, and VeLO\_No-Sweep can be deployed under interactions without any code changes, while epsilon-greedy was insufficient for multiple particle control. QP proved to be the most efficient and reached 100% success on two and three interacting particles. Its per-step displacement matrix implicitly absorbs the acoustic interaction force. Two algorithmic improvements were identified:
First, the tile-coded Q-learning with epsilon-decay, alpha-decay, and a shaped reward achieves 100% success at three particles with and without interactions.
Second, a modified Hessian regularisation for the VeLO is introduced in this work, together with the VeLO\_No-Sweep variant, which has no sweep phase.
These modified versions of the controllers are reliable across the tested range, with a success rate between (90-100%) at particle count of 1, 2, 3, with and without interactions. VeLO\_No-Sweep matches and sometimes exceeds standard VeLO at all conditions.
Q-learning matches QP's success, but at a far greater computational cost. The long training time (roughly three hours per table in the three particle case) and the inflexibility of the learned policy make QP the more practical choice for deployment in simulations.
A scalability stress test at n_P = 10 and n_P = 20, with targets arranged on a circle, further demonstrates that QP converges successfully under conditions involving particle counts far greater than three, including fully interaction-enabled dynamics.

Popular Abstract

Steering cells with sound: smarter algorithms for contactless particle control

When we think about moving individual cells, we tend to imagine tiny motorized nanorobots performing the task. In fact, one of the most advanced techniques for cell manipulation requires no moving parts. Instead, it relies on carefully controlled sound waves known as acoustophoresis. Making it work reliably for multiple cells at once is the problem this study tackles.

Every year, thousands of patients undergo procedures that require isolating specific cells from blood: circulating tumour cells for cancer diagnosis, stem cells for therapy, and immune cells for immunotherapy. Current methods are often slow, imprecise, or might damage the cells. Acoustophoresis offers a gentler alternative, where a sound wave is sent through a tiny water-filled chip, and cells drift toward the quiet regions of the wave. By rapidly switching between different wave patterns, a control algorithm can push particles toward any chosen destination without any physical contact.

Previous research showed that this works well for a single cell, but real clinical samples contain many cells at once. When several particles share the same chip, each cell radiates its own tiny acoustic force that pushes its neighbours, and every moving cell drags a wave of fluid that affects the others. These ‘’particle--particle interactions'' had been left out of earlier control algorithms.

To find out whether those interactions actually matter, five different algorithms were tested on the same set of initial conditions, run for two and three particles at a time, with and without interactions.

The algorithms span a wide range of approaches. The most mathematically precise, a quadratic optimiser called Perfect-Information Local Optimisation (PILOt), probes the chip's response at every time step and solves an equation to find the best blend of wave patterns. An online learner called Vector-based Local Optimisation (VeLO) instead watches how the particles actually move and updates its model on the fly.

The epsilon-greedy bandit algorithm balances exploitation and exploration. It usually chooses the acoustic wave pattern known to work best, but occasionally takes a random action on an untried frequency to discover an even faster path. Finally, a more complex reinforcement learning approach, Q-learning, expands on this concept by training offline and learning from observations.

The results were surprising in several ways. The quadratic optimiser actually improved when particle-to-particle interactions were introduced to the system. Because it probes the chip's behaviour at every time step, it captures the interaction features between cells and uses them as an extra steering force. The no-sweep variant of the online learner also performed well. It skips the calibration survey, which the standard VeLO starts with. VeLO\_No-Sweep learns entirely from the cells' actual movements, including how they affect each other, never falling behind the standard version. Q-learning kept pace with the quadratic optimiser in step count and matched its success rate. Q-learning's main limitation is that each table takes about three hours to train and cannot adapt once learned. The $\varepsilon$-greedy algorithm was unable to reliably handle two cells simultaneously with an acceptable success rate.

These results show which algorithms can cope with particle–particle interactions, and point to better lab-on-a-chip designs for precision medicine.

• author: Zowezer, Mohamad
• supervisor: Thierry Baasch; Alexander Edthofer; Shuo Huang; Guilherme Junqueira Perticarari
• organization: Division for Biomedical Engineering
• alternative title: Akustisk Partikelstyrning med Partikel-Partikel Interaktioner
• course: BMEM05 20261
• year: 2026
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• language: English
• additional info: 2026-20
• id: 9246017
• date added to LUP: 2026-07-02 13:38:31
• date last changed: 2026-07-02 13:42:57