Early detection of pore clogging in microfluidic systems with 3D convolutional neural network

In this study, we investigate whether the clogging phenomenon in a particulate suspension can be predicted from earlier observations of the system. Our research focuses on a microfluidic model system of polystyrene particles dispersed in a glycerol solution, where the onset of clogging can be controlled by adjusting the solution viscosity and flow rate. The microfluidic system allows for optical observations of the flow channels, providing detailed information on how particles are deposited in the flow passage. Using data collected from this model system, we developed a predictive algorithm based on 3D convolutional neural networks (3D CNN) that estimates the probability of clogging onset in the future based on past video frames of the system. Our results show that the 3D CNN can accurately predict clogging even under experimental conditions not encountered during training. The 3D CNN model with a depth of 9 was able to detect clogging after just 25 min, even though the actual clogging occurred after 118 min. This performance is superior compared to the 2D CNN, which detected clogging in 35 min under the same conditions. The high predictive performance indicates that the evolution of particle positions in the early stages of flow contains the necessary information for predicting clogging onset. Our findings have practical implications for the possibility of data-driven predictive maintenance of flow systems.