Accelerating Hydrodynamic Fabrication of Microstructures using Deep Neural Networks

Manufacturing of microstructures using a microfluidic device is a largely empirical effort due to the multi-physical nature of the fabrication process. As such, models are desired that will predict microstructure performance characteristics (e.g., size, porosity, and stiffness) based on known inputs, such as sheath and core fluid flow rates. Potentially more useful is the prospect of inputting desired performance characteristics into a design model to extract appropriate manufacturing parameters. In this study, we demonstrate that deep neural networks (DNNs) trained with sparse datasets augmented by synthetic data can produce accurate predictive and design models. For our predictive model with known sheath and core flow rates and bath solution percentage, calculated solid microfiber dimensions are shown to be greater than 95% accurate, with porosity and Young’s modulus exhibiting greater than 90% accuracy for a majority of conditions. Likewise, the design model is able to recover sheath and core flow rates with 95% accuracy when provided values for microfiber dimensions, porosity, and Young’s modulus. As a result, DNN-based modeling of the microfiber fabrication process demonstrates high potential for reducing time to manufacture of microstructures with desired characteristics.

💡 Research Summary

The paper addresses the longstanding challenge of designing and predicting the outcomes of microfluidic‑based microfiber fabrication, a process that is inherently multi‑physical and traditionally relies on trial‑and‑error experimentation. The authors propose a data‑driven framework that leverages deep neural networks (DNNs) to (i) predict key performance metrics—fiber diameter, porosity, and Young’s modulus—from known manufacturing inputs (core flow rate, sheath flow rate, and bath solution concentration) and (ii) invert this relationship so that desired performance specifications can be translated back into the appropriate process parameters.

To overcome the scarcity of experimental data, the study combines a modest set of 120 real‑world measurements with a large synthetic dataset generated via coupled computational fluid dynamics (CFD) and solid‑mechanics simulations. The synthetic data span the full design space, ensuring that the DNNs are exposed to a diverse range of operating conditions. All continuous variables are normalized to the