Abstract
Traditional design of fluid-based thermal management systems (TMS) relies on heritage architectures or computationally expensive numerical optimization. This paper presents a machine learning framework that extracts generalizable design knowledge from optimization data. By mapping optimal configurations to a dimensionless feature space, the framework identifies near-optimal architectures in seconds with minimal performance loss, enabling the design of complex systems using small-scale "building blocks."
Introduction
Efficient thermal management is critical for high-power density applications such as electric vehicles, data centers, and aerospace systems. While numerical optimization, specifically Open Loop Optimal Control (OLOC), can identify optimal designs for specific scenarios, its computational cost prohibits exhaustive searches across all possible architectures and operating conditions. This work addresses the need for "design intuition"—a method to rapidly guide engineers toward optimal system configurations without repetitive, full-scale optimization.
Methodology
The proposed workflow utilizes graph-based modeling to represent TMS architectures, where nodes signify heat exchangers (CPHXs) and edges represent fluid paths. Using Latin Hypercube Sampling, the authors generated hundreds of heat load "disturbances." For each disturbance, a brute-force optimization was conducted across all possible graph configurations to maximize thermal endurance (the time before any component exceeds its temperature limit).
A key innovation is the use of dimensionless, normalized features ($D_i = d_i / \sum d_i$) to represent heat loads. This transformation makes the extracted knowledge scale-invariant. Machine learning classifiers, including K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and Multi-Layer Perceptrons (MLP), were then trained to map these normalized features to the optimal configuration ID.
Mapping the Design Space
The extraction of design knowledge is visualized by mapping optimal configurations into a feature domain. For a 3-node system, the 13 possible configurations are distributed across the 2D and 3D normalized heat load space.
As shown in the feature space mapping, distinct regions correspond to specific optimal architectures. Interestingly, many configurations perform nearly identically in overlapping regions, suggesting a "good enough" principle where misclassification results in negligible performance penalties—often less than 1% of the true global optimum.
Results and Scalability
The KNN classifier emerged as the top performer, achieving a test accuracy of approximately 78% for 3-node systems. Despite this, the predicted configurations consistently yielded objective values within 1% of the global optimum. This demonstrates that the framework effectively captures the underlying physics of the design space even when exact classification is missed.
Furthermore, the authors demonstrate a "building block" approach. Knowledge extracted from 3-node and 4-node systems can be applied to decompose and design much larger systems (e.g., 8-node or 13-node architectures). This modularity bypasses the combinatorial explosion that occurs when attempting to optimize large-scale graphs directly, allowing designers to estimate optimal configurations in seconds rather than days.
Conclusion
This machine learning framework provides a significant computational speedup for TMS design. By transforming raw optimization data into a structured feature space, engineers can leverage pre-trained models to identify high-performing architectures for novel requirements. Future work will explore the use of Genetic Algorithms for more efficient exploration of larger configuration spaces and the integration of closed-loop control strategies.
Summary updated