上海科技大学知识管理系统

Computational fluid dynamics (CFD) is nowadays an indispensable tool in various applications, in which the computation is often performed within a finite box-like domain. This setting could differ from the reality, e.g., in open space, where the domain boundary plays a crucial role in accurately representing the consistent physical behavior using limited computational resources. Despite being extensively studied, existing domain boundary treatment lacks sufficient accuracy and generality, making current applications inevitably require different models in different conditions, some with a large multi-resolution domain to resemble an open space, which, however, consumes higher computational costs than expected. It can also bring larger numerical errors to eventually influence the overall accuracy of predicting useful physical quantities. In this work, we propose a novel data-driven solution to model the domain boundary. By simplifying the model equation with learnable coefficient and source terms, which are realized through the observation and action of multiple reinforcement learning (RL) agents, our new domain boundary model can be effectively applied in various scenarios with a small domain, without any fine-tuning for a specific setup, thus enhancing the computational accuracy, while reducing the costs in certain cases. We verify our model by conducting a series of numerical experiments in open domain setting as an example, which is difficult to be handled with a unified boundary model. Different flow regimes are considered, from laminar flow to flows with vortices and turbulence, in which pressure waves may co-exist if the flow is compressible. Good agreement with the reference flow is observed, with comparisons to existing methods to highlight enhanced accuracy and generality. To the best of our knowledge, this is the first work trying to successfully tackle complex domain boundary problem in a unified manner. It not only establishes a novel way to deal with domain boundary in CFD, but also underscores a significant potential inherent in the application of data-driven approaches to address related problems in partial differential systems.