【作者】 金戈； 李世鹏；

【Author】 Jin Ge;Li Shipeng;Beijing Institute of Technology;

【机构】 北京理工大学；

【摘要】 基于数据驱动的传统机器学习方法在精确处理流体流场预估问题时往往面临关键信息数据量较小、机器学习过程收敛难度大以及预测结果容易违反真实物理规律等问题。本文首先基于一种融合物理信息的深度学习框架,以典型的非线性偏微分方程为例,使用少量边界信息与初始信息获得了较高精度的解,然后进一步将此深度学习框架拓展到二维Navier-Stokes偏微分方程组的求解过程中,实现了在数据稀疏的小数据条件下的流场预估,最后展现了此框架在其他流动问题中的应用潜力。更多还原

【Abstract】 When traditional data-driven machine learning methods accurately deal with the problem of fluid flow field prediction, they often face problems such as small amount of key information data,difficult convergence of machine learning process and easy violation of real physical laws of prediction results. Firstly, based on physics-informed neural networks(PINN), this paper takes a typical nonlinear partial differential equation as an example and uses a small amount of boundary information and initial information to obtain a relatively high-precision solution. Then, this deep learning framework is further extended to the solution process of two-dimensional Navier-Stokes partial differential equations. Finally,the potential of this framework in other flow problems is demonstrated.更多还原