Accelerating Unsteady Fluid Dynamic Simulations for Microdevices for Droplet Generation Using Singular Value Decomposition and Recurrent Neural Network

Abstract

Digital twins are expected to be applied in the chemical process industry as important technology for improving productivity and profits. A key requirement is a flow model that reproduces the internal flow of equipment. Computational fluid dynamics (CFD) provides high-fidelity field data and is widely used to analyze flow and mass transfer, but unsteady simulations are computationally expensive because the solution must be advanced with very small time steps. To integrate CFD into digital twins, this cost must be reduced. In this study, the reduced order modelling (ROM) of CFD using singular value decomposition and recurrent neural network was applied to a T-shaped microdevice for droplet generation, which is used in particle production and related applications. The resulting ROM accurately predicts flow fields and droplet size across operating conditions while achieving an approximately 3,600-fold speed-up relative to CFD. Novelty is the implementation of a CFD-trained ROM as a digital-twin virtual sensor for microdroplet devices. The method enables near-real-time droplet-size estimation with demonstrated accuracy and robustness.