TOPOLOGY OPTIMIZATION OF MULTIFUNCTIONAL THERMAL MANAGAMENT SYSTEMS FOR AEROSPACE APPLICATION
Open Access
- Author:
- Mekki, Bashir
- Graduate Program:
- Mechanical Engineering (PHD)
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- October 10, 2022
- Committee Members:
- Daniel Haworth, Professor in Charge/Director of Graduate Studies
Alexander Rattner, Major Field Member
Xiaofeng Liu, Outside Unit & Field Member
Stephen Lynch, Chair & Dissertation Advisor
Vikram Shyam, Special Member
Xiang Yang, Major Field Member - Keywords:
- Topology Optimization
Heat Exchangers
Additive Manufacturing
CFD
Machine Learning
Experimental Testing - Abstract:
- Researchers have employed various techniques to enhance heat exchanger (HEX) performance. They have applied geometrical enhancement by introducing extended surfaces such as wavy, louvered, perforated, and offset strip fins to increase the heat transfer. However, traditional manufacturing has limited the complexity of the enhanced geometry. Recently, advances in manufacturing techniques such as additive manufacturing (AM), in conjunction with significant advances in modeling tools, have opened wide possibilities for developing and fabricating more complex geometries that are not traditionally manufacturable. As a result, topology optimization (TO), typically used in structural optimization, has emerged as a powerful design tool for heat exchangers, and much work is needed to establish its application in this area. This dissertation utilizes genetic algorithm-based topology optimization coupled with computational fluid dynamic (CFD), emerging machine learning methods, and modern manufacturing technologies such as AM to develop freeform high-performance heat exchanger fins for critical applications such as aerospace. First, a two-dimensional (2D) and three-dimensional studies demonstrated the ability to successfully couple TO and CFD to generate organic, high performance designs. New optimized designs are produced through a series of genetic algorithm (GA) mutation and crossover processes and evaluated using the CFD software OpenFOAM. This process is repeated until the improvement over the baseline design has stopped changing in a significant way. The results have shown performance improvement levels relative to the baseline of up to 60% in the case of 2D and up to 21% in the case of the 3D study. Overall, the significance of these two studies is that they successfully validated the ability to couple GA and CFD to produce better performing fin designs with organic shapes that are otherwise not possible to produce using traditional design methods. CFD proved to be computationally expensive and time consuming in the optimization approach. For this reason, one of the studies in this dissertation explored the ability to utilize machine learning algorithms as surrogates for CFD. This is achieved by utilizing the previously generated data from the 2D study as training data for an artificial neural network (ANN) known as multilayer perceptron (MLP). First, an MLP network is used to approximate the designs’ 2D images to a useful number of parameters that are paired with the CFD mesh as input to flow prediction parameters. Second, another MLP network is devised to predicted flow parameters (pressure, velocity, temperature) using the input from the first network and the CFD data. The trained models are tested on a set of unseen data, and it was found that the evaluation time took about 29 seconds compared to over 12 minutes when using the CFD. The trained models are coupled with the GA in place of CFD, and further used to explore the influence of GA convergence criterion. Finally, the optimized designs are additively manufactured, and experimentally tested to validate thermal and hydraulic performance. Overall, the heat transfer from the experiment for the tested designs agrees with that from the CFD results. However, the roughness from the AM process led to experimental pressure drop much higher than that predicted by the CFD for all the designs. Because of this, the optimized designs’ experimental fitness function improvement over the baselines is below what is observed in the CFD. Despite that, analysis of the experimental testing results suggests that imposing a limit on minimum and maximum fin size and using a more sophisticated fitness function such as the ratio of j-factor to f-factor could enhance the optimized designs fitness function.