Computational Fluid Dynamics (CFD) plays a pivotal role in the comprehension and forecasting of intricate fluid flow phenomena across a multitude of engineering and scientific disciplines. FLOWGEN represents an approach aimed at advancing computational fluid dynamics (CFD) workflows. It enables the online training of modern learning models concurrently with the cogeneration of time-evolving turbulent flow data at realistic scales. Conventional CFD methodologies frequently necessitate the utilisation of prior simulations, which can give rise to difficulties in terms of scalability, memory storage, and computational efficiency. In contrast to traditional methodologies, FLOWGEN performs data generation and model training simultaneously, effectively overcoming the challenges associated with memory storage.
This study employs a lattice Boltzmann solver optimized for graphics processing units (GPUs) to generate real-time time-evolving turbulent flow data. Concurrently, deep learning models are trained in real time using the generated data to develop highly accurate CFD surrogates. FLOWGEN integrates physics-driven simulations with deep learning frameworks at high scales. The integration of deep learning with CFD simulations has the potential to accelerate design processes, optimise resource allocation, and enhance decision-making in industries that rely on atmospheric flow understanding.
The anticipated acceleration in computational efficiency will facilitate decision-making processes, enhance design optimisation and improve resource allocation, thereby heralding a paradigm shift for industries reliant on atmospheric flow understanding. In addition to its specific applications in the field of atmospheric flows, FLOWGEN also anticipates broader implications for CFD workflows. By focusing on atmospheric flows in urban environments, FLOWGEN anticipates making significant contributions across various engineering and scientific domains. By enabling the rapid generation of highly accurate, time-resolved 3D simulations, FLOWGEN aims to facilitate the development of effective data-driven CFD surrogates and accelerate the design processes for wind farms, urban infrastructure, and numerous other engineering setups.
The initial experiments conducted with FLOWGEN have yielded promising results, demonstrating the potential for the rapid generation of accurate, time-resolved 3D simulations. The outputs of this study, including open-source software contributions to the PhyDLL and lbmpy libraries, are intended to facilitate knowledge sharing and collaboration within the European research community.