Physics-Guided Synthetic CFD Data Generation and Explainable Deep Learning Models for Automated Flow Pattern Classification

Computational Fluid Dynamics (CFD) analysis traditionally depends on manual interpretation of complex flow patterns through methods which are both subjective and time-consuming and require extensive domain expertise. This research presents an innovative framework that synergizes synthetic physics-informed CFD data generation with explainable deep vision models to enable automated flow pattern classification. The study constructs a comprehensive synthetic dataset using mathematical models that replicate realistic fluid flow behaviors across three regimes: laminar flows (Re: 2000) characterized by sinusoidal functions yielding smooth parallel streamlines, turbulent flows (Re: 2000) modeled with multi-scale chaotic and stochastic components, and separated flows exhibiting recirculation zones with exponential decay properties. Being aligned with established fluid mechanics principles, this physics-informed approach facilitates controlled parameter adjustments. The framework utilizes ResNet-50 as convolutional neural networks (CNN) attaining a test accuracy of 93.83%, and ViT-Base vision transformers achieving a test accuracy of 99.33%, to interpret velocity and vorticity field visualizations for flow pattern classification, enhanced by the Explainable Artificial Intelligence (XAI) technique Grad-CAM, which provides visual explanations to ensure model reliability. This approach can significantly benefit aerodynamics by improving the prediction of airflow behavior around aircraft. Furthermore, it offers potential for real-time aerodynamic analysis, supporting the development of more efficient and safer aviation technologies.