HeatGen: A Guided Diffusion Framework for Multiphysics Heat Sink Design Optimization

TL;DR

This work tackles the challenge of designing heat sink geometries under multiphysics convection–diffusion coupling while simultaneously minimizing pressure drop and preventing overheating. It introduces HeatGen, a diffusion-based generative framework that uses gradient guidance from differentiable surrogate models to steer Bézier-fin geometries toward low-pressure, thermally feasible designs. Key contributions include first application of denoising diffusion probabilistic models to heat sink topology under fully coupled physics, demonstration of up to 10% reductions in pressure drop compared with CMA-ES, and robust CFD validation alongside high surrogate accuracy. The approach offers scalable, fast inference suitable for digital-twin environments and sets the stage for future extensions with reinforcement learning and manufacturing constraints.

Abstract

This study presents a generative optimization framework based on a guided denoising diffusion probabilistic model (DDPM) that leverages surrogate gradients to generate heat sink designs minimizing pressure drop while maintaining surface temperatures below a specified threshold. Geometries are represented using boundary representations of multiple fins, and a multi-fidelity approach is employed to generate training data. Using this dataset, along with vectors representing the boundary representation geometries, we train a denoising diffusion probabilistic model to generate heat sinks with characteristics consistent with those observed in the data. We train two different residual neural networks to predict the pressure drop and surface temperature for each geometry. We use the gradients of these surrogate models with respect to the design variables to guide the geometry generation process toward satisfying the low-pressure and surface temperature constraints. This inference-time guidance directs the generative process toward heat sink designs that not only prevent overheating but also achieve lower pressure drops compared to traditional optimization methods such as CMA-ES. In contrast to traditional black-box optimization approaches, our method is scalable, provided sufficient training data is available. Unlike traditional topology optimization methods, once the model is trained and the heat sink world model is saved, inference under new constraints (e.g., temperature) is computationally inexpensive and does not require retraining. Samples generated using the guided diffusion model achieve pressure drops up to 10 percent lower than the limits obtained by traditional black-box optimization methods. This work represents a step toward building a foundational generative model for electronics cooling.

Figures (21)

• Figure 1: A 3D schematic view of the problem along with the simplified pseudo-3D model. It represents the conversion from the full 3D model to two 2D models, with boundary conditions indicated.
• Figure 2: Architecture of the 1D U-Net for geometry reconstruction of fin arrays
• Figure 3: Illustration of the downsampling module consisting of two residual blocks followed by a convolution layer with stride 2.
• Figure 4: One Resblock with skip connections and time embedding
• Figure 5: Graphical model of the denoising diffusion probabilistic model (DDPM). Dashed arrows denote the forward diffusion process $q(x_t|x_{t-1})$, while solid arrows represent the learned reverse denoising process $p_\theta(x_{t-1}|x_t)$.