Deep learning for reduced order modelling of wall bounded, turbulent flows

仿真（和详细的实验）湍流流体可以被定义为“数据丰富，知识贫乏”的活动：由于物理现象的复杂性，直接从大量的观测数据中学习和提取新知识几乎是不可能的.本文将利用高保真模拟积累的数据，利用前馈神经网络建立有效的降阶模型.降阶后，模型是可用的，人们可以快速估计流场的变化，随着参数的变化。这可以用作强化学习算法的环境。

Simulation (and detailed experiments) of turbulent fluid may be definedas a "data-rich, knowledge-poor" activity: the complexity of underlyingphysics is so high that it is almost impossible to learn and extract newknowledge from a large amount of the observed data directly.This thesis will focus on data accumulated with high fidelity simulationsto build effective reduced order models set up using a feed forward neural network.Once a reduced order model is available, one can have rapid estimates of flow field changes as the parameters are changed.This can be used as an environment for reinforcement learning algorithms.