Machine Learning for Computation Fluid Dynamics

Kunal Dewangan

Assistant Professor

Department of Mechanical Engineering

Kalinga University, Naya Raipur, Chhattisgarh, India, 492101

kunal.dewangan@kalingauniversity.ac.in

Keywords: Machine Learning, Computational Fluid Dynamics, Turbulence Modeling, Reduced Order Modeling, Flow Control, Optimization, Uncertainty Quantification, Data-Driven Models, High-Fidelity Simulations, Numerical Methods.

Machine learning (ML) has emerged as a powerful tool in the field of computational fluid dynamics (CFD), offering new possibilities for improving the accuracy and efficiency of fluid flow simulations. ML techniques enable the development of data-driven models that can capture complex flow phenomena and make predictions without relying solely on traditional numerical methods. These models learn from large datasets generated from high-fidelity simulations or experimental measurements, allowing for more accurate turbulence modeling, reduced order modeling, flow control, and uncertainty quantification.

In turbulence modeling, ML algorithms have shown promise in capturing intricate turbulence dynamics that are challenging to simulate accurately using traditional models. By training on extensive datasets, ML models can learn and reproduce complex flow features, leading to improved predictions of turbulence quantities. This has the potential to enhance the accuracy of CFD simulations, particularly for turbulent flows in engineering applications [1,2,3].

Reduced order modeling is another area where ML has demonstrated its value. CFD simulations can be computationally demanding, especially for large-scale problems or parametric studies. ML techniques can be used to develop reduced order models that approximate the full CFD solution at a fraction of the computational cost. By learning the underlying physics from a limited set of high-fidelity simulations, these models enable engineers and researchers to perform rapid design iterations and optimization tasks efficiently [4].

Flow control and optimization in CFD can also benefit from ML techniques. By coupling ML algorithms with CFD simulations, researchers can explore large design spaces and identify optimal configurations. Reinforcement learning, genetic algorithms, and surrogate modeling are among the ML methods used to optimize flow systems, leading to improved performance and efficiency.

Uncertainty quantification is a critical aspect of CFD simulations, considering the uncertainties associated with measurement errors, model assumptions, and numerical approximations. ML algorithms offer a means to quantify and propagate uncertainties, providing probabilistic predictions and enabling informed decision-making. Bayesian machine learning approaches, in particular, provide a framework for estimating uncertainties and incorporating them into the simulation results, enhancing the reliability and robustness of CFD predictions.

References:

1.      Ling, J., Kurzawski, A., & Templeton, J. (2016). Reynolds averaged turbulence modelling using deep neural networks with embedded invariance. Journal of Fluid Mechanics, 807, 155-166.

2. Noack, B. R., & Morzyński, M. (2017). Reduced-order models for flow control. Annual Review of Fluid Mechanics, 49, 181-209.
3. Jin, X., Yu, D., Zhang, Z., & Cao, W. (2019). Review on applications of machine learning techniques in computational fluid dynamics. Applied Sciences, 9(6), 1140.
4. Wang, Q., Yan, W., & Karniadakis, G. E. (2020). Physics-informed generative adversarial networks for stochastic differential equations. Journal of Computational Physics, 406, 109165.