Reproduce “HeatGen: A Guided Diffusion Framework for Multiphysics Heat Sink Design Optimization” on OpenResearch

About this paper

Hadi Keramati, Morteza Sadeghi, Rajeev K. Jaiman

The HeatGen framework optimizes heat sink geometries by employing a guided denoising diffusion probabilistic model (DDPM) that minimizes pressure drop while adhering to maximum temperature constraints. This approach yields up to a 10% reduction in pressure drop compared to traditional black-box optimization methods, demonstrating efficient generation of high-performing designs.

Key insights

• Probabilistic diffusion models can be effectively guided by differentiable physics-based surrogate models for complex, multi-objective engineering design problems, offering a stable and scalable alternative to prior generative and optimization methods.
• The iterative denoising process, when steered by performance objectives derived from surrogate model gradients, can reliably transform random noise into physically consistent, high-performing thermo-fluidic geometries.
• A multi-fidelity data generation pipeline, coupled with an automated workflow, is crucial for building robust generative and surrogate models capable of accurately capturing complex non-linear thermo-fluidic behaviors.

Summary by alphaXiv.