Multi-task image-based deep learning for boiling analysis: Material recognition and heat flux prediction

Abstract

Pool boiling, a fundamental heat transfer process, has been extensively studied due to its importance in various industrial applications. This paper presents a multi-task deep learning model for simultaneous material recognition and heat flux quantification from boiling process images, providing a resource-efficient solution for engineering applications requiring precise thermal analysis. The proposed model utilizes a shared feature extraction backbone with attention-enhanced convolutional blocks and a multi-task output head to jointly handle material classification and heat flux regression tasks within a single framework. A weighted loss function is incorporated to balance the learning dynamics between tasks, enabling optimized performance for both material classification and heat flux quantification. Experimental results demonstrate the model's high accuracy in both tasks, with a material recognition accuracy of 100 % and a mean absolute error (MAE) of 0.094 W/cm² for heat flux prediction, underscoring its reliability for practical deployment in real-time, accurate thermal monitoring and analysis. Future work will explore integrating multi-modal data, such as acoustic data, to further improve predictive performance and broaden the model's applicability in complex thermal environments.