报告题目：Integrating Machine Learning Techniques in Turbulence Closure Models

报告时间：2024-11-16   10:00-11:00

报 告 人：连复桑  教授  （University of  Waterloo）

报告地点：老外楼301

Abstract:Industries such as aerospace, automotive, chemical, nuclear, hydroelectric, and wind power use numerical simulations of turbulent flows to design safe and efficient systems. However, the inherent complexity of turbulence equations makes these simulations computationally demanding. The widely-used Reynolds-averaged Navier-Stokes (RANS) approach simplifies these equations but often introduces significant errors due to the eddy viscosity approximation in turbulence closure models. Despite decades of research, existing turbulence models frequently fail to capture all relevant industrial physics. Recently, deep learning has emerged as a promising tool to enhance turbulence closure modeling. Unlike traditional methods based on intuition and heuristics, machine learning can derive complex functional relationships from data, potentially overcoming the limitations of conventional models. This seminar will explore various machine-learning approaches aimed at improving turbulence closure models. It will focus on an approach that corrects the Reynolds stress anisotropy tensor—a primary source of error in RANS models—by training a model to predict high-fidelity closure terms from low-fidelity inputs. The corrected anisotropy tensor is then integrated into the momentum equation, resulting in improved mean-field predictions. However, incorporating machine learning outputs into coupled partial differential equations demands careful attention to numerical stability and solution conditioning.