Study on Pneumatic Flow Control Algorithm Based on Physics-Informed Bayesian Optimization for Flow Cytometry

This study compared different flow-rate control algorithms based on vision-based indirect measurements in a pneumatic control system for polydimethylsiloxane microfluidic chips used in flow cytometry. Microfluidic systems are difficult to control precisely using a mathematical model alone owing to channel fabrication errors and bubble formation, and conventional iterative learning control (ILC) converges quickly, but is unstable under measurement noise. Thus, a physics-informed Bayesian optimization (PIBO) method that registered the Poiseuille law as the prior mean function of a Gaussian process was proposed. To emulate the experimental uncertainty, the measurement, offset, and spike noise were applied to a virtual chip model, and ILC, standard BO, and PIBO were compared. Under increasing noise, ILC exhibited persistent oscillation after convergence, whereas PIBO maintained stable convergence by statistically absorbing noise. PIBO was further applied to an actual pneumatic flow-control system, where it reliably converged to the target flow rate, thereby confirming its practical feasibility.