Session: Research Posters

Paper Number: 173012

Ai-Supported Multi-Objective Optimization of Microfluidic Channel Design for Saliva-Based Point-of-Care Diagnostic Devices 

Abstract

Introduction

Point-of-care (POC) diagnostic devices offer a noninvasive and accessible method for monitoring systemic health, particularly in underserved populations. Such devices must operate reliably outside controlled laboratory environments, creating engineering challenges in balancing performance, manufacturability, and production cost. Traditional design approaches, which rely heavily on iterative computational simulations and manual refinement, are time and resource intensive, creating bottlenecks in device development cycles. Recent advances in artificial intelligence (AI) and machine learning (ML) have shown strong potential in automating engineering tasks. 

To address this challenge, the current work explores the use of AI/ML to optimize microchannel design for saliva-based POC diagnostics. By leveraging surrogate modeling and multi-objective optimization, we aim to reduce the computational burden of repeated high-fidelity simulations and capture domain knowledge to be used for reliable design decisions. 

Contribution of the work

This research investigates how AI/ML can support and accelerate the channel design of microfluidic POC diagnostic devices by addressing the following questions: (1) Can simulation-informed AI models uncover interpretable parameter relationships in microfluidic channel design? (2) Can this approach reduce reliance on trial-and-error optimization by providing a principled understanding of design spaces? The main contributions of this work include: (1) Developing a surrogate modeling workflow that uses COMSOL-generated simulation data to train a neural network and integrates NSGA-II for multi-objective optimization in microchannel design. (2) Demonstrating that simulation-informed AI models can uncover interpretable parameter relationships, embedding engineering domain knowledge into data-driven design and enabling more explainable and efficient decision-making.

Methodology

The microfluidic channel in this study was designed based on inertial microfluidics to passively induce Dean flow, a secondary flow field that occurs in curved pipes or channels. A spiral microchannel geometry was chosen for its dual functionality to utilize Dean flow for both mixing and large particle separation. This geometry was parameterized and simulated using COMSOL Multiphysics to generate simulation data. A parametric sweep was performed across six geometric features, including channel width, height, and the number of turns in the spiral. The  average Dean number and pressure drop across the channel served as performance objectives for surrogate modeling and optimization. 

To uncover the relationships among design parameters and performance metrics, a neural network surrogate model was developed on the simulation dataset. Multi-output regression-based feature selection was applied to reduce input dimensionality, identify the most significant geometric parameters, and enhance model interpretability. Multi-objective optimization was performed using the NSGA-II algorithm to identify design candidates that simultaneously minimize pressure drop and maximize Dean number. This workflow enables efficient exploration of the design space without repeated high-fidelity simulations, while also uncovering which parameters most significantly influence performance.

Preliminary results and conclusions

Preliminary results demonstrate potential of the AI-assisted framework to improve microfluidic channel design efficiency and predictive accuracy. The surrogate model was evaluated 10 times before and after applying regression-based feature reduction to ensure statistical robustness. The mean percent error of Dean number predictions significantly decreased from 5.81% (STDEV = 3.12) to 3.04% (STDEV = 2.29). While the error of pressure predictions increased from 7.41% (STDEV = 8.61) to 9.04% (STDEV = 4.02), the decrease in standard deviation suggests the model is more stable. Pearson correlation between surrogate predictions and COMSOL simulation outputs improved from r = 0.917 to 0.981 for Dean number, and from r = 0.895 to 0.964 for pressure. Paired t-tests confirmed the improvements were statistically significant (p < 0.05).  These results indicate that removing noisy or irrelevant features improves both correlation and statistical agreement with simulation data. This work contributes an interpretable framework for accelerating microfluidic channel design while embedding domain knowledge into AI-guided workflows. Future work will leverage these modeling results to expedite physical prototyping using the CADworks3D ProFluidics 285D BioMed printer.

Presenting Author: Ronald Jose San Jose State University

Presenting Author Biography: Ronald Jose is pursuing an M.S. degree in mechanical engineering at San Jose State University (SJSU), San Jose,
California. He earned his B.S. degree in mechanical engineering at the University of California, Los Angeles (UCLA) in 2023.

Authors:

Ronald Jose San Jose State University
Yunjian Qiu San Jose State University
Lin Jiang San Jose State University
Yun Wang San Jose State University