Generative Thermal Design Through Boundary Representation and Deep Reinforcement Learning

Loading...
Thumbnail Image

Date

2022-09-08

Authors

Keramati, Hadi

Advisor

Hamdullahpur, Feridun

Journal Title

Journal ISSN

Volume Title

Publisher

University of Waterloo

Abstract

Advancement in Additive Manufacturing (AM) allows fabrication of complex geometries. This provides the opportunity to think beyond hand-designed topologies. The main goal of this dissertation is to build a Generative Design (GD) framework which generates several optimal topologies. In this framework, Machine Learning (ML) accelerated predictive models in conjunction with a precise geometry representation for thermal design automation are used to reach an effective design space exploration. A size optimization algorithm is presented to optimize the geometry of Manifold-microchannel heat exchanger with straight fins using genetic algorithm and a hybrid Computational Fluid Dynamics (CFD) model. Challenges for process development pertinent to this design configuration using AM are also introduced particularly in microchannel part which can be addressed by efficient design methods. Reinforcement Learning is a branch of machine learning with strong capabilities in sequential decision making to maximize a reward. This decision making tool along with a parametric approach for geometry representation are utilized for fin shape optimization in heat exchanger design. Machine learning models and neural networks provide predictive tools for a wide range of applications. In this thesis, Convolutional Neural Networks (CNNs) are deployed to successfully predict heat transfer and pressure drop of the geometries that can efficiently explore the design space. A high level accuracy is seen in predicting direct CFD results from shapes saved as images. Online computing for optimization is complex and computationally expensive. The pre-trained CNN models are used as an alternative for computational engine in GD. This method drastically reduces the time required for one episode of reinforcement learning from several minutes to few seconds. The design space is expanded to multiple fin shape optimization using Multi-Agent Proximal Policy Optimization (MAPPO) in which each shape is controlled in a decentralized way while value learning is performed in a centralized way. We validate our method using Multi-agent Particle-world Environment (MPE) for high dimensional action space. It is shown that cooperative interaction of agents with a shared reward results in optimal thermal design solutions with reduced pressure drop and enhanced heat transfer. A robust generative thermal design framework is developed with which there is no need for discretized design domain or derivation with respect to the domain settings. Proposed method provides the tools and knowledge for efficient use of time, physical space and computational resources particularly for microchannel heat exchanger design.

Description

Keywords

Deep Learning, Multi-agent Reinforcement Learning, Geometry Processing, Heat Transfer

LC Keywords

Citation