Breadcrumb
- Home
- Publications
- Proceedings
- 2025 AIChE Annual Meeting
- Engineering Sciences and Fundamentals
- Data Science in Flow & Transport Phenomena
- (474c) Intelligent Modeling of Multiphase Microdispersed Processes
In this work, we proposed the intelligent modeling framework that integrates machine learning (ML) with microfluidic experiments to enable accurate prediction, mechanism interpretation, and active optimization of operation conditions.
We introduce an innovative closed-loop workflow empowered by ML for exploring bubble flow characteristics efficiently and demonstrate it in constricted microchannels as a case study. The workflow incorporates: (1) an advanced convolutional neural network that facilitates microscopic image processing for bubble statistics; (2) a random forest model associated with the Bayesian exploration strategy that optimizes design of experiments by targeting at uncertainties in the parameter space; (3) model explanation techniques that enlighten insights into bubble breakup mechanisms from the data-driven perspective. Furthermore, reverse optimization is exemplified to prepare target bubble swarms in the constricted microchannel by utilizing established ML models. This work underpins the prospective role of collaborative AI in microfluidics, significantly reducing the required workload for applications or mechanism studies.
To further promote the role of AI in mechanistic understanding, the integrated "tree-based ML + SHAP" framework emerges as the data-mechanism dual-driven tool to offer a global perspective from extensive experiments in fundamental research. We compile a comprehensive dataset comprising over 1800 droplet-based experiments using 39 capillary-embedded microchannels and 31 biphasic fluid systems. Four ensemble tree-based ML models represent advances over the prior models in improving predictive performance. Notably, the XGBoost model achieves the best predictive performance with the highest test R2 of 0.988 and the lowest test RMSE of 0.027. Importantly, the models have the potential to test the droplet generation performance in the absence of experiments due to the generalization across unseen flow conditions, channel geometries, and biphasic systems, providing guidance in designing microchannels or operating conditions. The main factors influencing the droplet generation are found consistent with previous expertise knowledge, confirming the capability of models to capture fundamental physical behavior rather than merely fitting data. Therefore, some previously underexplored insights are revealed to open promising avenues for future research.
By combining these two efforts, our study presents a scalable, interpretable, and data-efficient ML framework for modeling and optimizing complex multiphase microfluidic systems, bridging predictive modeling with physical understanding and paving the way for intelligent, self-optimizing microfluidic platforms.