Enhancing Microparticle Separation Efficiency in Acoustofluidic Chips via Machine Learning and Numerical Modeling

An integrated approach for enhancing microparticle separation efficiency in acoustofluidic lab-on-a-chip systems is presented, combining numerical modeling in COMSOL 6.2 Multiphysics^® with reinforcement learning techniques implemented in Python 3.10.14. The proposed method addresses the limitations of traditional parameter tuning, which is time-consuming and computationally intensive. A simulation framework based on LiveLink™ for COMSOL-Python integration enables the automatic generation, execution, and evaluation of particle separation scenarios. Reinforcement learning algorithms, trained on both successful and failed experiments, are employed to optimize control parameters such as flow velocity and acoustic frequency. Experimental data from over 100 numerical simulations were used to train a neural network, which demonstrated the ability to accurately predict and improve sorting efficiency. The results confirm that incorporating failed outcomes into the reward structure significantly improves learning convergence and model accuracy. This work contributes to the development of intelligent microfluidic systems capable of autonomous adaptation and optimization for biomedical and analytical applications, such as label-free separation of microplastics from biological fluids, selective sorting of soot and ash particles for environmental monitoring, and high-precision manipulation of cells or extracellular vesicles for diagnostic assays.