Topology optimization of cooling channels based on swarm intelligence

Kun Wang, Tao Wang, Xianglin LiStructural 2025 · DOI

TL;DR

This paper introduces a novel approach to designing cooling channels by treating fluid particles as "intelligent" agents that naturally seek out heat. By using swarm intelligence instead of traditional, computationally heavy gradient-based optimization, the researchers achieved comparable cooling performance with a 22x speedup in calculation time.

What Problem Does This Paper Address?

In high-power electronics and aerospace engineering, getting heat out of a system is a life-or-death matter for the hardware. Engineers use Topology Optimization (TO) to find the most efficient layout for internal cooling channels. However, traditional TO for "multi-physics" problems (where fluid dynamics and heat transfer are coupled) is a nightmare.

Standard methods like the "variable density method" require massive computational resources, complex mesh partitioning, and often struggle to converge when the geometry gets complicated. The authors identify a "computational wall" where the time required for optimization becomes prohibitively long for practical engineering applications. They sought a "cost-effective" alternative that simplifies the physics without sacrificing too much accuracy.

Methods

The researchers moved away from traditional Finite Element Analysis (FEA) and instead used NetLogo, a multi-agent simulation platform. They modeled the system using two types of agents:

  1. Immobile Agents ("Patches"): These represent the solid material (e.g., a silicon plate). They "know" their temperature and conduct heat to their neighbors based on Fourier’s Law.
  2. Mobile Agents ("Coldbugs"): These represent the cooling fluid. They move through the solid domain, absorbing heat via simulated convection.

The optimization happens through two distinct agent-based models:

  • Chain-type Flow Model: Agents move in queues (like streamlines). They follow "Heat-seeking" rules to find hotspots and "Aggregation" rules to cluster together, which reduces fluid resistance. This is ideal for designing parallel channel structures.
  • Grid-based Flow Model: Agents move through a pre-defined grid. They balance a desire to find heat with an "Exit Rule" (a field function that pulls them toward the outlet). This naturally produces complex, tree-like branching structures.

To validate their results, the authors compared their NetLogo simulations against STAR-CCM+, a gold-standard commercial CFD (Computational Fluid Dynamics) software.

Datasets

As this is a physical simulation and optimization paper rather than a machine learning paper, it does not use traditional training datasets. Instead, the models were evaluated using:

  • Physical Parameters: Based on standard properties for Silicon (solid) and Water (fluid).
  • Validation Benchmarks: A 49-point temperature grid was used to compare the accuracy of the swarm intelligence model against the high-fidelity STAR-CCM+ numerical results.

Key Insights and Ideas

  • Implicit Optimization: Unlike traditional methods that require an explicit global objective function (like "minimize peak temperature"), this method relies on local rules. Optimization emerges naturally from the balance between an agent's "instinct" to find heat and its strategy to reach the exit efficiently.
  • Simultaneous Physics and Design: In this model, the simulation is the optimization. As the "coldbugs" find their paths, they are effectively drawing the final channel architecture in real-time.
  • The "Aha" Moment on Speed: The most striking insight is the efficiency gain. By simplifying the fluid-structure interaction into agent behaviors, the researchers bypassed the need for the iterative, heavy matrix inversions typical of CFD-based optimization.

Key Results

  • Speedup: The agent-based method completed the optimization in ~7 minutes on a standard laptop. In contrast, the gradient-based method took 154.7 minutes on a high-end workstation (24 cores, 128GB RAM).
  • Accuracy: Despite the simplifications, the accuracy was remarkably high. The Chain-type model showed an average accuracy of 95.4% compared to commercial software, while the more complex Grid-based model achieved 91.2%.
  • Thermal Performance: The optimized grid-based channels successfully reduced the average plate temperature from an initial 70°C down to approximately 60°C, with peak temperatures converging around 71°C, demonstrating effective hotspot mitigation.

Limitations and Future Work

While promising, the authors acknowledge several hurdles:

  • 2D vs. 3D: The current study is limited to two-dimensional plates. Real-world cooling systems often require complex 3D manifolds.
  • Energy Consumption: The rules for minimizing fluid pressure drop (pumping power) are currently simplified. Future versions need more robust "cost" functions for the agents to better simulate energy loss.
  • Branching Logic: While the grid model produces tree-like structures, the mechanisms for fluid confluence (merging) and branching could be refined to better reflect actual fluid behavior.

This research suggests that for early-stage design exploration, "intelligent" particles might be a much faster way to find the right shape than traditional, "brute-force" math.

Summary updated

Abstract

This paper proposes a swarm intelligence based structural topology optimization method on cooling channels. This method employs the swarm intelligence concept to simulate thermal conduction phenomena in solid domain and heat exchange phenomena between fluid and solid domain. Research indicates that for cooling structures with chain-type channels, this method achieves optimization of intervals between channels; for cooling structures exhibiting grid-based channels, this method allows for optimized distribution of flow in different channels. The method was implemented on NetLogo, and the feasibility and accuracy of the simulation model were verified by commercial numerical software. This study show a prominent cost-effectiveness and new approach for topology optimization of internal cooling channels.
BibTeX
@inproceedings{wang2025topology,
  author = {Kun Wang and Tao Wang and Xianglin Li},
  title = {{Topology optimization of cooling channels based on swarm intelligence}},
  booktitle = {Structural},
  year = {2025},
  doi = {10.1007/s00158-025-04076-x},
}
Figure