A data driven framework for optimizing droplet microfluidics with residual block and Fourier enhanced networks.

Droplet-based microfluidic devices enable the generation of uniform droplets with precise control over size and production rate-key factors in diagnostics and pharmaceutical screening. Achieving such control is essential for enhancing efficiency and reducing operational costs. However, traditional design approaches rely heavily on complex physical models or labor-intensive trial-and-error processes, making them time-consuming and expensive. In this study, we present a data-driven framework that employs machine learning to predict droplet size and generation frequency, while simultaneously optimizing device geometry. Our novel methodology integrates forward prediction of droplet characteristics with inverse design for geometric ratio optimization. Droplet formation dynamics within a co-flow microfluidic device were simulated using the Lattice Boltzmann method, generating a comprehensive dataset of 658 cases across both dripping and jetting regimes. We developed two innovative machine learning models: the Fourier-Enhanced Network (FEN), which utilizes Fourier series to decompose input features into harmonic components; and the Residual Block Network (ResBNet), which incorporates skip connections within residual layers to capture complex nonlinear patterns. ResBNet demonstrated high accuracy in predicting relative droplet radius, Strouhal number, and optimized geometric ratios, while FEN offered computational efficiency and robust performance across a broad range of flow rates. To facilitate practical applications, we developed DesignFlow-an open-source platform that automates the design and optimization of microfluidic devices. DesignFlow significantly reduces simulation time and cost, enabling rapid prototyping for biomedical and pharmaceutical applications (github.com/AlirezaSamari/DesignFlow).