Physics-informed Machine Learning for Turbulent Separated Flows Using Smart Sensors

Real-time prediction of incoming turbulence has important practical utility in the control and monitoring of fluid systems. Flow estimation using smart sensing techniques will improve our ability to predict and control these flows. However, current sensor-based flow reconstruction methods have not been tested to effectively resolve multi-scale, chaotic and unsteady turbulent phenomena that are commonly encountered in industry. Moreover, the relatively more powerful machine learning (ML) methods still require large amounts of data to train these models on real turbulent flows, rendering their potential to be used in deployed systems impractical. The proposed 24-week multi-disciplinary project involves the development of novel physics-informed ML methods that use sparsely placed wall pressure sensors to predict entire three-dimensional unsteady turbulent flowfields. The success of the project will lead to a robust ML-driven framework that (1) requires significantly less training data, (2) yields more physically consistent predictions and, (3) can predict missing dimensions from limited information.

Faculty Supervisor:

Robert Martinuzzi

Student:

Partner:

Université de Poitiers

Discipline:

Engineering

Sector:

Education

University:

University of Calgary

Program: