Inverse design of microfluidic concentration gradient generator using deep learning and physics-based component model

摘要

This paper presents a new paradigm of deep neural network (DNN) for the inverse design of microfluidic concentration gradi-ent generators (μCGGs) with complex network topology. In this method, a concentration gradient (CG) and design parameters yielding the CG are, respectively, used as inputs and outputs of DNN, and the relationship between them is mapped. Several new elements are also proposed, including utilization of fast-running physics-based component model in the closed form to generate a large amount of data for DNN learning which otherwise is not available through computationally demanding computational fluid dynamics (CFD) simulation; and a divide-and-conquer strategy and DNN formulation combining clas-sification and regression to mitigate many-to-one design complications for enhanced accuracy. Several DNN structures are investigated and developed, including single fully connected neural network (FCNN), convolutional neural network, and a new cascade FCNN for a divide-and-conquer implementation. Case studies are performed on a triple-Y μCGG to evaluate design performance of the proposed method in a six-dimensional space that only includes sample concentrations at inlet reservoirs as design parameters, and in a nine-dimensional design space, to which inlet flow pressures are also added. It is verified in high-fidelity CFD simulation that widely used CGs can be produced using DNN-predicted design parameters accurately with average error ＜4% and ＜ 8.5% relative to the prescribed CGs, respectively, in the six- and nine-dimensional design space. The learned design rules are packaged into the DNN model that allows to generate accurate μCGGs designs instantaneously (～3 ms) and eliminates requirements of simulation and optimization knowledge, facilitating distribution of the design capabilities to microfluidic end users.