(425e) Predicting the Traffic of Confined Microfluidic Droplets Using Machine Learning
AIChE Annual Meeting
Thursday, November 19, 2020 - 9:00am to 9:15am
Current approaches to predict droplet decisions and behaviors are based on resistive network models where droplets are treated as point resistors. Although simple and elegant, these models do not incorporate local event dynamics at a junction and therefore do not reproduce experimentally observed complex behaviors. In this study, we investigate the capability of machine learning to predict the traffic of droplets in a loop with the objective of identifying decision-rules governing complex behaviors that are hitherto difficult to conceive from fluid physics models or human-guided observations. To apply machine learning, we perform experiments with trains of droplets entering the loop at different spacings, enabling us to capture images of thousands of droplets making decisions at the junction. These image data sets are used as training sets to develop the learning models.
As a starting point, we encode droplet decisions into a logistic regression model. For the case of large droplet spacings in the train, logistic regression successfully predicts the outcome based on the number of droplets in each branch. However, as the droplet spacing in the train decreases, logistic regression gradually fails to predict the outcomes because of local event dynamics. For the smallest spacing, the prediction accuracy was found to fall below 60%. Additionally, the logistic regression model trained on a given spacing was found to be not capable of generalization to the other unseen spacings which limits its suitability for predicting droplet traffic.
Next, we used convolutional neural network (CNN) to predict droplet decisions in the loop. Surprisingly, CNN is capable of learning and predicting the droplet decisions with high accuracy regardless of the spacing. In contrast to logistic regression, CNN is also capable of generalization to new unseen situations. To investigate the decision-making process of the neural network, we use an occlusion mask that continuously occludes different parts of the input image while we monitor the change in classification score providing insights on the features that ultimately make the network to make its decision. Our analysis revealed that the neural network is not only sensitive to the number of droplets in the branches, but also to their positions in branches and at the junction. Stated differently, the network itself learned the importance of event dynamics at the junction. Our results suggest that deep learning models have significant potential to predict the traffic of drops and other deformable particles in even more complex networks than studied here. Their success in predicting complex transport phenomena might uncover essential elements that need to be incorporated into fluid physics models.