Advancing Fluid-Based Thermal Management Systems Design: Leveraging Graph Neural Networks for Graph Regression and Efficient Enumeration Reduction

Saeid Bayat, Nastaran Shahmansouri, Satya R. T. Peddada, Alex Tessier, Adrian Butscher, James T. AllisonarXiv 2023 · DOI · arXiv

Abstract

Designing optimal thermal management systems (TMS) is a complex task due to the combinatorial explosion of potential component configurations. Traditional high-fidelity simulations, such as Open-Loop Optimal Control (OLOC), are computationally expensive, limiting the exhaustive exploration of the design space. This paper proposes a Graph Neural Network (GNN) framework to predict system performance directly from graph-based representations of TMS architectures. By performing graph regression, the model achieves high ranking accuracy (Kendall Tau > 0.97), enabling a reduction in required simulations by 92% to 99.8% while successfully identifying optimal designs. The approach demonstrates strong generalization from simpler to more complex system topologies, offering a scalable path for automated engineering design.

Introduction

In modern engineering, the design of thermal management systems (TMS) often relies on "design heritage"—incremental improvements on existing architectures. While safe, this approach frequently misses radical, high-performance configurations. To find truly optimal designs, researchers employ enumeration, where every possible combination of pumps, valves, and heat exchangers is modeled and simulated.

However, as the number of components increases, the design space grows exponentially. Evaluating each architecture requires solving complex OLOC problems to determine maximum thermal endurance. These high-fidelity simulations are the primary bottleneck, often requiring days of computation for even moderately sized systems. This work introduces a GNN-based surrogate model to bypass this bottleneck by predicting performance directly from system topology.

Methodology

Graph Representation and Data Generation

Thermal systems are naturally modeled as directed graphs. Nodes represent physical components like tanks, junctions, and cold plate heat exchangers (CPHXs), while edges represent the fluid flow paths.

Physical thermal management system (a) and its corresponding graph representation (b) used for GNN training.

To train the model, a synthetic dataset of 37,329 TMS instances was generated. A small subset (approximately 30%) was evaluated using a traditional OLOC solver to provide ground-truth performance labels (thermal endurance time).

Graph Attention Network (GAT) for Regression

The core of the proposed method is a Graph Attention Network. Unlike standard neural networks, GNNs are designed to process non-Euclidean data, making them ideal for learning how component connectivity impacts thermal performance. The model performs graph regression to predict a continuous performance value. The workflow involves training on the labeled subset and then using the GNN to rank the remaining 70% of the design space. Engineers then only need to run expensive simulations on the top-ranked candidates.

Key Insights: Topology and Complexity

A critical finding of this research is that system performance is heavily dictated by topology rather than just component count. The GNN successfully learned that under varying heat loads, specific serial or parallel configurations are superior.

Examples of different graph topologies for systems with varying numbers of components, illustrating the combinatorial complexity.

Furthermore, the model exhibited impressive generalization. A GNN trained on simpler systems (3–5 nodes) could accurately predict the performance of 6-node systems it had never encountered, suggesting that the learned topological features are robust across scales.

Results and Discussion

The GNN framework demonstrated exceptional efficiency and accuracy:

  • Computational Efficiency: In a case study of 4,051 designs, the GNN identified the true optimal architecture by evaluating only the top 7 candidates, a 99.8% reduction in simulation effort. On average, the method reduced simulation requirements by over 92%.
  • Ranking Accuracy: The model achieved a Kendall Tau rank distance of 0.97 to 0.99. This high correlation indicates that even if absolute error exists, the relative ranking of designs is nearly perfect, which is the primary requirement for design optimization.
  • Inference Speed: While full OLOC evaluation for the 6-node case took 12 hours on a 64-core workstation, the GNN provided predictions for the entire set in seconds.

Conclusion and Future Work

This work proves that GNNs can significantly accelerate the design of complex fluid-based systems by learning the underlying physics of connectivity. Future research will focus on handling transient heat loads through recurrent GNN architectures and exploring "Type 2" problems where the design space is too vast to even enumerate, necessitating generative graph models.

Summary updated

BibTeX
@misc{bayat2023advancing,
  author = {Saeid Bayat and Nastaran Shahmansouri and Satya R. T. Peddada and Alex Tessier and Adrian Butscher and James T. Allison},
  title = {{Advancing Fluid-Based Thermal Management Systems Design: Leveraging Graph Neural Networks for Graph Regression and Efficient Enumeration Reduction}},
  year = {2023},
  eprint = {2311.14874},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2311.14874},
}
Figure