Evolutionary design of conductive pathways using a generative autoencoder

To enhance performance and energy efficiency in compact electronic systems, effective thermal management through miniaturized heat sinks is essential. This study introduces a generative design framework for optimizing bi-material heat sink architectures suitable for space-constrained electronics. Conventional thermal optimization methods often rely on gradient-based techniques or parametric geometric shapes, which limit adaptability and design resolution. In contrast, our approach integrates topology optimization, constructal theory, and bio-inspired principles via a generative model that couples a variational autoencoder with evolutionary algorithms. Trained on a diverse dataset of microleaf structures, constructal patterns, and synthetic conductive networks, the model addresses the area-to-point conduction challenge, a common scenario where heat must be efficiently transferred from a broad source area to a narrow sink. Compared to standard methods, the proposed design pipeline reduces average device temperatures by 10% and peak hotspot temperatures by up to 1%. Additionally, it achieves a 13% increase in Pareto front hypervolume, indicating superior trade-offs among conflicting thermal objectives. By enabling smooth feature transitions and adaptive geometries through latent space interpolation, the method supports modular and scalable design of unit cells and full assemblies. A demonstrative case assembles these cells into a manufacturable three-dimensional heat sink tailored for spatially varying thermal loads. The methodology can extend beyond conductive heat sinks to encompass conjugate heat transfer and multi-physics systems by adjusting design objectives or finite element formulations. This work offers a blueprint for generative thermofluids design in applications such as phase-change materials, heat pipes, and cold plate systems.