Machine learning for predicting microfluidic droplet generation properties

Abstract

Computational Fluid Dynamics (CFD) simulators are an integral tool for designing commercial microfluidic systems — allowing engineers to estimate key performance metrics prior to physical prototyping. However, CFD simulations can take from hours to weeks, depending on the simulation fidelity. This introduces a significant delay in the design iteration loop. In this work, we employ machine learning methodologies to provide estimates of performance metrics directly from microfluidic geometries in inkjet printheads — providing approximate performance metrics in seconds. Specifically, we develop a Convolutional Neural Network (CNN) based approach that operates directly on voxelized slices of printhead geometry to predict characteristics of droplet generation. This approach does not encode significant biases for the task or physics knowledge and relies on a training set of 20,000 prior simulation results to learn. Despite this, our experiments on this large dataset demonstrate that the learned model can closely approximate the CFD results at a fraction of the time cost for some performance metrics — opening the doors to real-time metric estimation as part of the microfluidic system design process. Further, we examine the learned latent representation and find it encodes a reasonable notion of geometric similarity between printhead architectures. This can allow engineers to search for existing designs with similar characteristics and help reduce duplicated effort.