Generative constructal design for thermal flow systems
File(s)
Author(s)
Ignuta-Ciuncanu, Matei-Cristian
Type
Thesis
Abstract
This thesis advances the field of thermal design by unifying evolutionary principles with generative methods. Rooted in the Constructal Law, the theory that systems evolve to facilitate easier access to what flows, the work proposes a novel methodology for the design of thermal systems that are free to morph, adapt, and optimize across spatial and temporal scales. This freedom is expressed through a multi-level generative framework that combines evolutionary search, variational autoencoders, and finite element multi-physics solvers to create function-driven flow architectures. Unlike biomimicry which is descriptive and fractal or parametric methods which are prescriptive, the method described in this thesis is predictive; it reveals not just efficient configurations of thermal systems, but how they evolve with new physics‑driven demands.
Across four classes of thermal systems, conductive heat sinks, hierarchical thermal metamaterials, waste heat recovery, and electrochemical cooling, the designs evolved in silico demonstrate a persistent tendency toward improved flow access and reduced resistance. For area-to-point conduction, generative designs reduced multi-objective performance by up to 13% over state-of-the-art gradient-based optimization methods while achieving 6x faster convergence. In hierarchical metamaterials, the generative framework produced multi-objective thermal cloaks in under 5 minutes on a standard workstation, demonstrating rapid convergence and adaptability to complex performance goals. Experimental validation in pulsating convection channels revealed up to 35% enhancement in heat transfer using realistic valve profiles under low-frequency, high-amplitude conditions. In battery cooling systems, evolutionary vascular networks reduced hotspot non-uniformity by up to 37% and pumping power by 50% compared to straight-channel parametric cold-plate designs. Notably, all the designs presented are achievable on a low-budget consumer computer, highlighting the accessibility of the approach.
The work establishes the generative constructal paradigm as a scalable design philosophy for thermal systems---one that is both informed by nature (carbon intelligence) and enabled by machine learning (silicon intelligence).
Across four classes of thermal systems, conductive heat sinks, hierarchical thermal metamaterials, waste heat recovery, and electrochemical cooling, the designs evolved in silico demonstrate a persistent tendency toward improved flow access and reduced resistance. For area-to-point conduction, generative designs reduced multi-objective performance by up to 13% over state-of-the-art gradient-based optimization methods while achieving 6x faster convergence. In hierarchical metamaterials, the generative framework produced multi-objective thermal cloaks in under 5 minutes on a standard workstation, demonstrating rapid convergence and adaptability to complex performance goals. Experimental validation in pulsating convection channels revealed up to 35% enhancement in heat transfer using realistic valve profiles under low-frequency, high-amplitude conditions. In battery cooling systems, evolutionary vascular networks reduced hotspot non-uniformity by up to 37% and pumping power by 50% compared to straight-channel parametric cold-plate designs. Notably, all the designs presented are achievable on a low-budget consumer computer, highlighting the accessibility of the approach.
The work establishes the generative constructal paradigm as a scalable design philosophy for thermal systems---one that is both informed by nature (carbon intelligence) and enabled by machine learning (silicon intelligence).
Version
Open Access
Date Issued
2025-12-19
Date Awarded
2026-03-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Martinez-Botas, Ricardo
Wang, Li-Liang
Publisher Department
Department of Mechanical Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)