Thermal fluid fields reconstruction for nanofluids convection based on physics-informed deep learning.

Area of Science:

• Computational fluid dynamics
• Deep learning applications in physics
• Heat transfer and fluid mechanics

Background:

• Traditional methods for thermal fluid field reconstruction often rely on mesh-based approaches, limiting flexibility.
• Physics-informed neural networks (PINNs) offer a promising avenue for integrating physical laws into deep learning models.
• Accurate prediction of performance characteristics (e.g., Nusselt number, Fanning friction factor) is crucial for engineering design.

Purpose of the Study:

• To propose a physics-informed deep learning model for accurate thermal fluid field reconstruction.
• To enable the calculation of key performance characteristics from reconstructed fields.
• To investigate the model's performance against convolutional neural network-based methods and classical machine learning techniques.

Main Methods:

• Development of a deep learning model using fully-connected layers to map design variables and spatial coordinates to physical fields.
• Integration of conservation laws into the loss function to enhance physical interpretability and model accuracy.
• Validation using forced convection data of water-Al2O3 nanofluids, including analysis of training data size and extrapolation capabilities.

Main Results:

• The proposed deep neural network accurately reconstructs thermal fluid fields and predicts performance characteristics.
• The model demonstrates superior predictive performance for performance characteristics compared to classical machine learning methods.
• The method shows no constraints on mesh generation, offering greater flexibility than traditional approaches.

Conclusions:

• Physics-informed deep learning provides an effective and interpretable approach for thermal fluid field reconstruction.
• The developed model accurately predicts key performance metrics, outperforming existing machine learning techniques.
• The study highlights the potential of PINNs for complex fluid dynamics problems and explores crucial aspects like extrapolation performance.