AI-Driven Prediction and Optimization of Heat Transfer in Thermal Energy Systems

Abstract

Accurate prediction and optimization of heat transfer remain central challenges in the design of thermal energy systems. Traditional approaches such as computational fluid dynamics (CFD), while reliable, are often computationally expensive and unsuitable for real-time applications. This study explores the potential of machine learning (ML) as a fast and reliable alternative. A hybrid dataset of ~12,000 points was developed from experiments and CFD simulations, covering a wide range of Reynolds numbers, Prandtl numbers, and geometrical configurations. Four ML models-Random Forest, XGBoost, Artificial Neural Networks (ANN), and Physics-Informed Neural Networks (PINNs)-were trained and compared. ANN and XGBoost delivered the highest accuracy (R² > 0.95, RMSE < 5%), while PINNs provided physically consistent predictions by embedding governing equations into the learning process. Compared with CFD, the AI models reduced computation time by more than 95%, achieving near real-time predictions. To extend beyond prediction, optimization techniques were integrated into the framework. Genetic algorithms improved heat exchanger fin geometry, resulting in an ~18% increase in heat transfer coefficient with only a marginal rise in pressure drop. Reinforcement learning optimized PCM storage operation, achieving ~14% higher efficiency. Experimental validation, supported by uncertainty analysis (±3-4%), confirmed that AI predictions align closely with both CFD and measurements. The findings demonstrate that AI can act as both a predictive surrogate and an optimization engine for thermal systems. Looking ahead, such frameworks hold strong potential for integration into digital twins, enabling continuous monitoring, adaptive control, and autonomous operation in renewable energy, industrial heat recovery, and electric vehicle thermal management.