SMART-Flow: Enhancing Fluid Dynamics Simulations with Machine Learning for Scalable Model Reduction and Accurate Prediction

Aims: To improve the efficiency of fluid dynamics simulations by applying model reduction and prediction techniques through machine learning, ultimately reducing computational costs while maintaining near-perfect accuracy. Background: Fluid dynamics simulations are critical in various engineering and physical science applications. However, their computational complexity often leads to significant time and resource consumption. To address this issue, integrating computational physics with machine learning- based methods offers promising opportunities for performance enhancements. Objective: To develop a machine learning-driven model reduction approach that significantly reduces computational costs while preserving simulation accuracy for fluid dynamics studies. Method: The approach combines Principal Component Analysis (PCA) and Proper Orthogonal Decomposition (POD) for initial model reduction. Physics-informed Convolutional Neural Networks (CNNs) output signals are compressed using deep learning-based autoencoders, extracting the key features of fluid dynamics. Machine learning models, including regression methods and neural networks, are trained on this reduced dataset to predict fluid behavior under various conditions. The framework was implemented using Python, with TensorFlow and PyTorch for machine learning and OpenFOAM and ANSYS Fluent for conventional computational fluid dynamics (CFD). Parallel computing was employed to maximize resource utilization during large-scale simulations. Result: The integrated framework demonstrated significant improvements in computational efficiency and maintained accuracy when validated against high-fidelity CFD simulations and experimental data. Conclusion: The proposed machine learning-based model reduction approach effectively enhances the accessibility and practicality of fluid dynamics modeling, offering substantial computational gains and maintaining accuracy, thus broadening its application and appeal. The proposed framework also holds potential for patentability, given its innovative integration of machine learning and fluid dynamics modeling.