Can Graph Neural Networks Help Identify Promising Thermal Management System Architectures Among Vast Numbers of Possibilities?

Abstract The more efficient a thermal management system must be across a diverse range of conditions, the greater the intricacy of its design is required. Yet, the computational bottlenecks of enumerating and analyzing potential system architectures often render the best solutions impractical. This study evaluates the feasibility of accelerating system architecture performance evaluation by leveraging graph neural network (GNN)-based regression as a cost-effective alternative to traditional open-loop optimal control (OLOC) analysis. We examine a case study where enumerating all possible system architectures is computationally feasible, but directly analyzing the performance of each is not. Instead, we analyze a small subset of architectures and use these data to train a generalizable surrogate model capable of evaluating the remaining architectures within the available computational budget. After the training, the predicted performance values are sorted to obtain the estimated best configurations. Then, there are two options: (1) select the highest-ranked configuration as the optimal solution or (2) choose a small subset of test data with the highest estimated ranks and evaluate them using OLOC to get a more accurate result for the correct optimal configuration. Our results show that training a GNN with approximately 30% of the architectures was sufficient to predict the performance of the remaining 70% with an average mean squared error (MSE) loss of 0.6, achieving a 92% reduction in computational cost overall.