摘要
Efficient control of microrobot-assisted microfluidic mixing requires accurate prediction of spatiotemporal flow evolution and identification of optimal actuation strategies. However, these objectives remain challenging due to inherently nonlinear fluid–structure interactions and the experimental effort required for repeated measurements. In this work, a time-conditioned generative modeling framework was developed to enable the prediction and optimization of microrobot-assisted microfluidic mixing directly from experimental image data. Specifically, a conditional generative adversarial network (cGAN) was trained using mixing image sequences obtained from a single actuation location of the microrobot. Subsequently, a rotation-equivariant transformation strategy was incorporated to extend prediction capability to multiple spatial locations without requiring additional experimental data. Through this approach, the temporal evolution of mixing performance was accurately reproduced, and close agreement with experimental observations was achieved. The trained model was further integrated with a Bayesian optimization framework to identify optimal actuation time allocations of the microrobot across multiple spatial locations. As a result, a non-uniform actuation strategy was observed to enhance the mixing efficiency significantly, achieving 95.12% within 240s. The proposed framework enabled physics-informed, data-driven prediction and optimization of flow-induced mixing and established a scalable approach for spatiotemporal control in microfluidic systems.
[Display omitted]
•A data-driven conditional generative adversarial network (cGAN) framework was developed for microrobot-assisted micromixing prediction.•A rotation-equivariant approach was employed and integrated into the framework for multi-location prediction.•The Bayesian optimization method was applied to identify an optimal actuation strategy.•A mixing efficiency of 95.12% was achieved within a limited actuation time.•Time-efficient locations were identified under placement constraints.