Session: 02-05-01: Data Driven Design

Paper Number: 146073

146073 - Active Machine Learning Empowered Digital Twin Modeling for Heat Sink Design 

As electric vehicles (EVs) and plug-in hybrid vehicles (PHEVs) become increasingly prevalent, efficient thermal management of power inverters and converters is crucial for safe operation and prolonged lifespan. The design process of heat sink often necessitates an optimization phase that can be computationally intensive due to the repeated, intricate analyses of the system. This paper addresses the challenges in designing and manufacturing effective cooling systems for these components, introducing a cutting-edge methodology that leverages active machine learning to empower digital twin modeling for heat sink design. Traditional methods involve time-consuming physical prototyping and disconnected design-manufacturing processes, leading to prolonged development cycles and high costs. To overcome these challenges, a novel digital twin platform is proposed, integrating physics-based modeling with advanced data-driven techniques. The platform utilizes a multiphysics simulation model for constructing the physical twin and employs an active machine learning (AML) approach based on Gaussian process (GP) modeling for the digital twin. Our approach begins with offline experiments and extends to an online AML framework, facilitating real-time interaction with the physics model for deployment in production settings. By strategically leveraging labeled samples, our method outperforms random data point selection, resulting in a notable enhancement in surrogate modeling efficacy. This innovative approach offers a more efficient system performance evaluation and automatic uncertainty quantification, guiding the exploration of optimal heat sink designs while reducing energy consumption and development costs. The paper discusses the proposed method, presents a numerical example, and concludes with insights into the efficacy of the approach for enhancing thermal management in electric vehicle systems.

In conclusion, the active machine learning empowered digital twin modeling approach presented in this study offers a promising solution for enhancing the efficiency and effectiveness of heat sink design optimization. By integrating a multiphysics model with a Gaussian process-based surrogate model and employing active learning techniques, the computational burden associated with complex thermal analyses is significantly reduced. The iterative selection of data points for model training leads to improved accuracy and faster convergence to optimal design solutions. This innovative framework not only streamlines the design process but also enables real-time interaction with the physics model, facilitating informed decision-making and cost-effective product development in the realm of thermal management for advanced engineering applications。 In future endeavors, our focus will extend towards integrating and accounting for the complexities of heat sink geometry. This entails incorporating intricate microchannel designs and leveraging topology optimization principles to enhance performance. Additionally, we aim to delve into bio-inspired design methodologies, drawing inspiration from nature to further innovate in heat sink design. These advancements promise to elevate our model's capability to simulate and optimize heat dissipation systems, paving the way for more efficient and sustainable thermal management solutions.

 

Presenting Author: Yanwen Xu University of Texas at Dallas

Presenting Author Biography: Yanwen Xu is an assistant professor of the Department of Mechanical Engineering at the University of Texas at Dallas. Yanwen earned her Ph.D. in industrial engineering from the University of Illinois at Urbana-Champaign, and B.S. in Mathematics and Applied Mathematics from Huazhong University of Science and Technology in China. Her research interests lie in the field of engineering design for reliability analysis and system optimization, and prognostics and health management, where she focuses on developing new machine learning methods and computational tools to improve the reliability of engineered systems.

Authors:

Jacob Coaty University of Texas at Dallas
Hao Wu Purdue University
Majid Minary University of Texas at Dallas
Yanwen Xu University of Texas at Dallas