Enhancing heat transfer coefficient predictions in complex geometries through hybrid machine learning approaches

• Published in: Thermal science and engineering progress 2024-10, Vol.55, p.103017, Article 103017
• Main Authors: Kalpana, V., Jessy Sujana, G., Thyagarajan, K., Lalitha, R.V.S., Talasila, Vamsidhar, Mohan Jadhav, Makarand
• Format: Article
• Language: English
• Subjects: Complex Geometries; Computational Fluid Dynamics (CFD); Convolutional Neural Networks (CNNs); Data-driven Modeling; Hybrid Machine Learning; Multiphysics Modeling; Physics-Informed Neural Networks (PINNs); Thermal Engineering
• ISSN: 2451-9049
• DOI: 10.1016/j.tsep.2024.103017

•Novel hybrid CNN-PINN model for heat transfer coefficient prediction.•Superior accuracy compared to traditional and pure ML methods.•Effective generalization to unseen complex geometries.•Significantly faster than high-fidelity CFD simulations.•Potential for real-time thermal management applications. Precise prediction of the heat transfer coefficients over complex geometries is critical for bypassing expensive and time-consuming trial-and-error approaches in designing thermal systems in electronics, automotive, aeronautics and maritime industries. Therefore, the main goal of this study was to propose a novel hybrid machine learning model that combines the power of convolutional neural networks (CNNs) for learning geometric features with physics-informed neural networks (PINNs) for enforcing physical constraints outperforms both empirical correlations and purely absolute error of 52K and a relative test dataset. The hybrid approach displays superior generalisation capabilities in predicting the heat transfer coefficients of unobserved complex geometries in real-world scenarios, such as wavy-walled channels, pin–fin heat sinks and helical coil heat exchangers. At the same time, the hybrid model displays extraordinary computational efficiency as the prediction times are orders of magnitude lower than those required by high-fidelity CFD simulations. We have contributed a powerful tool towards rapid design iteration and optimisation in thermal engineering, which could foster a revolution in the design of heat exchangers, electronic cooling strategies and general real-time thermal management systems for various industrial applications.