Title : Improving heat exchangers by manipulating wall-bounded flow guided by machine learning algrithms
The PhD will be co-supervised by : Prof. Datta Gaitonde ( Ohio State University) & Dr. Lionel Agostini (Pprime institute - CNRS)
The first half of the Phd will be carried out at Ohio State university and the second half at Poitiers.
Keywords : flow control, heat transfer, wall-bounded flow, thermal boundary layer, numerical simulations, reduced order model, machine learning, data-driven algorithms, deep reinforcement learningKeywords : flow control, heat transfer, wall-bounded flow, thermal boundary layer, numerical simulations, reduced order model, machine learning, data-driven algorithms, deep reinforcement learning
Objectives and Scientific Challenges :
A goal of this research project is to propose solutions to improve the performance of heat exchangers by using a combination of active control strategies to maximise heat-transfer capacity along the interface fluid/structure and minimise the pressure loss due to friction between the fluid and the wall. Efforts to meet these constraints hinge on careful attention to the mixing processes in the wall-bounded flow and more specifically to the turbulence conditions. The proposed near-wall control combines plasma actuators to generate mobile large streamwise vortices with rapid oscillatory wall motion in the spanwise direction to enhance mixing produced by the relatively small streaky structures occurring at the edge of the viscous sublayer ( Agostini et al. 2014,2017 and 2021 ). High fidelity computational investigations, including Direct Numerical Simulations (DNS) as well as Large eddy simulation (LES), will be performed of the combined flow-control methodologies, and aided by the application of Machine Learning algorithms (ML), in an integrated fashion to maximize heat transfer and minimize frictional loss. The ML algorithms encompass, respectively, dimensionality reduction and low-dimensional dynamical modelling (Agostini 2020), and optimal control law design. Various techniques will be used to integrate physical knowledge into ML algorithms. Since they are physics- informed, rapid convergence towards more robust solutions is achieved. Initially, studies will be undertaken at relatively low Reynolds numbers, where control strategies may be more easily identified. Knowledge gained at low-Reynolds numbers will then be leveraged to extend control strategies for higher Reynolds-numbers flows.
This project is interdisciplinary, with both fundamental science and practical engineering significance. It combines and advances the cutting edge of several advanced scientific concepts and capabilities; as such, it has the potential for substantial academic impact in broader contexts, regardless of ultimate practical use. First, the exploitation of advanced DNS for combined drag and heat-transfer optimisation, subject to oscillatory wall motion is novel and is likely to advance DNS into a new flow-control direction of interest to the flow-physics community. Second, the adaptation of ML to optimise thermo-fluids systems is in its early infancy; the application of this emerging technology for optimising drag and heat transfer in the manner proposed has many unique aspects. The research outputs are expected to help model, predict and control heated wall-bounded flows. In addition to major contributions to our knowledge of near-wall physics and heat transfer, this work has important practical implications on harvesting and transporting energy. The programme will result in influential publications and possibly patents.
Requirements:
The candidate should have a Master's degree in Fluid Mechanics/Applied Mathematics/Machine Learning and have an appetite for interdisciplinary work and machine learning.
Key Duties and Responsibilities:
This project provides the opportunity to overcome several timely challenges. The first objective is to define a low-dimensional dynamic model of the wall- bounded flow by combining different machine learning algorithms, such as the auto-encoder combined with cluster-based reduced order modeling, CROM, and Dynamic Mode Decomposition, DMD.
Once the ROM is developed, deep reinforcement learning (RL) will be used to identify an optimal forcing law that modulates heat transfer.
The position requires collaboration in a multidisciplinary research environment comprised of mathematicians, computer scientists and engineers.
Specific responsibilities include:
The Pprime laboratory is a CNRS Research Unit. Its scientific activity covers a wide spectrum from materials physics to mechanical engineering, including fluid mechanics, thermics and combustion. During his/her stay in France the PhD student will be attached to the team Curiosity. The laboratory is classified as a ZRR (Zone à Régime Restrictif), which means that all recruitment is subject to prior authorisation by the Defence Security Officer.
Offer RequirementsEngineering: Master Degree or equivalent
Computer science: Master Degree or equivalent
Physics: Master Degree or equivalent
ENGLISH: Excellent
Contact Information