Impact of Dataset Size and Hyperparameters Tuning in a Variational Autoencoder for Structure–Property Mapping in Porous Metamaterials

ABSTRACT This study extends a deep learning framework for the inverse reconstruction of open‐cell porous metamaterials targeting specific hydraulic properties such as porosity and intrinsic permeability. Building on our recent work that employed a property‐variational autoencoder (pVAE) for structure–property mapping, the current contribution examines the impact of hyperparameters tuning, including latent space dimension size and network architecture, on model performance. Additionally, the effect of dataset size on the quality of the trained model is systematically analyzed. The framework uses synthetic 3D porous microstructures with unidirectional flow. The pVAE incorporates a convolutional encoder–decoder architecture with a Gaussian latent space. Training minimizes a composite loss that includes reconstruction loss for data fidelity, Kullback–Leibler divergence for latent space regularization, and regression loss to align with target hydraulic properties. The proposed latent space enables efficient generation of structures tailored to specific properties, supporting scalable, data‐driven metamaterial design.