Explainable machine learning-based prediction of mixing efficiency in passive microfluidic micromixers using experimental data

Abstract

In microfluidic systems, passive micromixers play a crucial role by ensuring fluid is well-mixed in microscale channels without the need for external energy input. However, there is a significant challenge in the design of passive microfluidic devices to obtain efficient mixing under laminar flow. An explainable machine learning framework was designed to predict mixing efficiency in passive microfluidic micromixers, using experimentally obtained data. The experimental data from the zigzag and split-flow micromixer configuration were used to assess the effect of flow rate, Reynolds number, viscosity and design on mixing performance. Three regression-based machine learning models, namely Linear Regression, Decision Tree Regressor and Random Forest Regressor, were developed and tested with Root Mean Square Error (RMSE) and coefficient of determination (R²) metrics. The model with the best predictive performance was the Linear Regression model, which had a value of RMSE equal to 0.0207 and R² equal to 0.8827, showing good agreement between the mixing index values predicted and experimental. Explainable artificial intelligence analysis using SHAP and feature-importance evaluation revealed that flow rate and Reynolds number were the dominant factors influencing mixing efficiency, while viscosity exhibited a negative effect on mixing performance. The results show the viability of combining microfluidic data from experiments with explainable machine learning methods to analyze and optimize passive micromixers. The proposed framework offers a meaningful and efficient solution to the problem of repeated experiments and CFD-based optimization in microfluidic system design.