# ML-Driven optimization of two-phase microfluidic cooling using acoustofluidic bubble actuation and nanoarray-coated micropin structures

## TL;DR

This paper introduces a new microfluidic cooling method using sound-driven bubbles and nano-coated structures, optimized with machine learning to improve heat transfer efficiency.

## Contribution

The novel integration of acoustofluidic bubble actuation, nanoarray-coated micropins, and machine learning for two-phase cooling optimization.

## Key findings

- LSTM models outperformed DNNs in predicting thermal performance with lower prediction errors.
- Initial temperature and chipset material were identified as the most influential factors affecting heat transfer.
- Acoustofluidic excitation was found to be the primary positive contributor to thermal performance.

## Abstract

This study presents a novel two-phase microfluidic cooling strategy integrating acoustofluidic bubble activation, nanoarray-coated micropin structures, and machine learning-guided optimization. Unlike conventional passive cooling, ultrasound-driven bubble actuation stabilizes boiling, prevents drying, and ensures uniform heat distribution, while nanoarray-coated micropins enhance liquid refilling via capillary forces and improve surface wettability. Experimental data were analyzed using deep neural networks (DNN), long short-term memory (LSTM) models, and statistical correlation methods (Spearman and Kendall), with interpretability provided by SHAP (DeepSHAP) and partial dependence plots (PDP). A comparison of model performance revealed that the LSTM achieved lower prediction errors than the DNN across all evaluated metrics (MAE 0.0055, SMAPE (0.8), and RMSE 0.0072), indicating its outperforming performance. SHAP and statistical analyses identified initial temperature as the most influential factor affecting heat transfer coefficient (HTC), followed by chipset material (S30-120 and stainless steel). Secondary factors include chipset configuration (SS, S30-120, S-nanorod, S-nanosheet) and nanoparticle type (SiO₂, ZnO), which significantly modulate bubble dynamics and thermal performance. PDP results highlight acoustofluidic excitation as the primary positive contributor, while flow rate and nanostructured surfaces provide moderate enhancements.

The online version contains supplementary material available at 10.1038/s41598-025-23871-6.