When a linear train of moderately confined droplets is introduced at high feeding frequency, drop collisions inevitably occur at the microfluidic junction. These collisions alter the fate of the droplet arriving at the junction, and counter-intuitively make them enter the long branch with higher hydraulic resistance. We study this phenomenon by tracking thousands of droplets, analyzing their trajectories, and choices they make at the junction. Using this dataset, we engineer three features that encode the angle during the collision, and develop a logistic regression model capable of predicting the decision rules for droplets colliding at a junction. We argue that this system could be a testbed for developing machine learning models to predict traffic of deformable particles in complex networks where collisions dominate.
