2019/02/28 by Arvind Mohan, Mohan, Arvind, Don Daniel +6 · 2 citations
Computer Science · Physics and Astronomy · Engineering · #Generative Adversarial Networks and Image Synthesis #Model Reduction and Neural Networks #Fluid Dynamics and Turbulent Flows
paper · pdf · doi:10.48550/arxiv.1903.00033
High-fidelity modeling of turbulent flows is one of the major challenges in\ncomputational physics, with diverse applications in engineering, earth sciences\nand astrophysics, among many others. The rising popularity of high-fidelity\ncomputational fluid dynamics (CFD) techniques like direct numerical simulation\n(DNS) and large eddy simulation (LES) have made significant inroads into the\nproblem. However, they remain out of reach for many practical three-dimensional\nflows characterized by extremely large domains and transient phenomena.\nTherefore designing efficient and accurate data-driven generative approaches to\nmodel turbulence is a necessity. We propose a novel training approach for\ndimensionality reduction and spatio-temporal modeling of the three-dimensional\ndynamics of turbulence using a combination of Convolutional autoencoder and the\nConvolutional LSTM neural networks. The quality of the emulated turbulent\nfields is assessed with rigorous physics-based statistical tests, instead of\nvisual assessments. The results show significant promise in the training\nmethodology to generate physically consistent turbulent flows at a small\nfraction of the computing resources required for DNS.\n