Year: 2024
Type: conference-paper
Abstract: Efficient heat management in the context of semiconductor miniaturization demands unconventional solutions to navigate constrained design spaces. The emergence of unsupervised learning as a versatile design companion capable of processing multi-disciplinary datasets presents promising avenues for streamlined bio-inspired solutions for fundamental conduction problems. As a case study, we tackle the pivotal area-to-point problem that led to the formulation of the Constructal Law. This study proposes a novel methodology harnessing Denoising Diffusion Probabilistic Models (DDPM) to seamlessly integrate features from topology optimization, constructal theory, and biomimetic structures. We use DDPM as a digital non-equilibrium complex system to generate geometric patterns for conductive heat sink design. These patterns are sampled to address the original formulation of the area-to-point problem. This study emphasizes the significance of morphological freedom in addressing multi-objective problems, reinforcing arguments first formulated within the constructal paradigm. Our framework facilitates low-cost sampling of intricate shapes capable of serving multiple temperature objectives by synthesizing principles from different fields. We show the emergence of dendritic structures to solve distribution problems in an unsupervised learning scenario, drawing a parallel between information and energy flows. This research underscores the transformative potential of Generative AI in blending design features across disparate disciplines, a potent tool for developing conductive heat sink solutions beyond deterministic optimization approaches.
DOI: 10.59277/clc.2024.09
Institution: Imperial College London
Language: English
Subfield: Mechanical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4407045665