Microfluidic Injector Models Based On Neural Networks

Abstract

A functional modeling technique is developed for components of a microfluidic system and applied to three common injector topologies. This technique uses sparse numerical simulation to train a neural network to provide compact, explicit, and accurate component models. The resulting models are compatible with analytic system simulation environments, making complex design synthesis and optimization feasible, unlike standard techniques using computationally expensive numerical simulation. The neural network models are accurate to numerical simulation with mean squared errors less than 10 -4 . In explicit form, the neural network models are very fast taking less than 1s per evaluation.