Abstract Spinodal topologies (STs), inspired by material phase transition phenomena, exhibit complex geometries generated by Gaussian random fields (GRFs). These dual-phase structures feature smooth transitions and low-curvature designs, offering enhanced mechanical response compared to traditional surface- or strut-based metamaterials. Their flexible design spaces allow both periodic and non-periodic topologies, making them promising candidates for multiphysics applications. This work presents a computationally efficient data-driven framework for optimizing the mechanical and thermal properties of STs. A convolutional neural network (CNN) is trained on homogenized elastic and thermal properties obtained from finite element analyses. Next, the CNN model is coupled with an optimization solver to design dual-phase STs to improve multiphysics objectives. The inherent variations of STs introduced by the GRF formulation limit the ability of gradient-based optimizers to efficiently search for the design space. To overcome this challenge, a gradient-free global optimization algorithm is integrated into the surrogate model to find optimum designs with superior mechanical and thermal performance.
Yıldız et al. (Mon,) studied this question.