Physics-Informed Machine Learning for Enhanced Thermal-Fluid Predictions in Complex Engineering Systems: A Hybrid CFD-Data Driven Framework

Abstract

Computational fluid dynamics has become an indispensable tool for analysing complex thermal-fluid systems in engineering applications ranging from particle detector cooling to electronics thermal management. However, the computational cost of high-fidelity CFD simulations limits their use in real-time monitoring, design optimization, and uncertainty quantification. This investigation presents a physics-informed machine learning framework that combines the predictive accuracy of CFD with the computational efficiency of data-driven models. The proposed approach uses a K-nearest neighbours algorithm constrained by the governing conservation equations to reconstruct temperature and velocity fields from sparse sensor measurements. The framework is validated against three-dimensional CFD simulations of a simplified enclosure representative of particle detector thermal management systems. Results demonstrate that the physics-informed KNN model achieves a mean absolute error of 2.34 K for temperature predictions and 0.087 m/s for velocity magnitude, representing a 67% reduction in error compared to pure data-driven interpolation. The model successfully captures thermal stratification patterns and buoyancy-driven flow structures with a computational speedup of approximately 850 times compared to the full CFD solution. Sensitivity analysis reveals that the optimal number of neighbours varies between 3 and 7 depending on the local flow regime, with lower values performing better in regions of high gradient. The framework provides a practical methodology for developing fast surrogate models that preserve physical consistency while enabling real-time thermal monitoring and control in complex engineering systems.