Generative Constructal Design of a Multi-Physics Heat Sink for Managing Transient Thermal Loads

This research applies generative constructal design to develop a multi-material heat sink for transient thermal loads. High conductivity material (HCM) is used in conjunction with Phase Change Material (PCM) to provide fast heat transfer pathways for efficient cooling and thermal stabilisation during power dissipation spikes. The anisotropic design domain containing HCM and PCM is modelled using the finite elements. The anisotropic design domain is modeled using the Finite Element method. A modified variable-viscosity model interpolates thermal properties across HCM, PCM – liquid and solid phases. The system solves parametric Navier-Stokes equations with the Boussinesq approximation, coupled with a transient diffusion equation including an enthalpy source term. The physics solver, implemented in Python using FEniCS, captures natural convection inside the PCM. This high-fidelity model enables the generative design agent to optimize both conduction and free convection pathways. The HCM configuration is generated using a variational autoencoder (VAE) trained on synthetic and bio-inspired geometries optimized for multiple flow systems. The generative model was trained and tested on a dataset with synthetic and bio-inspired features that have undergone evolutionary optimisation for multiple flow systems. As a result, the learned genetic content can be transferred to new physics, with most iterations focused on fine-tuning to the specific thermal loading conditions of the problem. This approach enhances the access of the model to the physics of the system compared to lumped apparent heat capacity model that represent the system as solely conductive. Moreover, we show that the application of constructal theory on digital information flows yields results with enhanced freedom-to-morph that can adapt to the requirements of the system. The results of this exploration show that the structures emerging from generative design methods exhibit a strong dependence on model fidelity, highlighting the necessity of multi-physics finite element formulations for high-efficiency integrated solutions. The methodology developed can be replicated for multi-physics systems to automate constructal and evolutionary design.