An end-to-end generative pipeline for designing Triply-Periodic Minimal Surface scaffolds

Triply periodic minimal surface (TPMS) scaffolds are architectures for bone tissue engineering. However, inverse scaffold design remains challenging, since multiple design parameters can yield similar target properties, and conventional trial-and-error approaches are computationally expensive. This study tackles the inverse problem by proposing an end-to-end generative artificial intelligence pipeline for inverse TPMS scaffold design able to provide a diverse pool of scaffold candidates. The proposed method also controls for scaffolds diversity, while ensuring a tolerance of 5% on all target properties. Three conditional generative adversarial networks were developed for Gyroid, Diamond, and Schoen’s Wrapped Package (IWP) architectures to generate TPMS parameters from five target features, two representing morphologic features and three mechanical properties. A feedforward neural network trained to address the direct problem was used as surrogate regressor to guide generators training and evaluate generated designs. Scaffolds diversity was accounted for in the generators loss function, using a covariance-based metric of the generated TPMS parameters for a given set of target features. The regressors achieved high accuracy on the test set, with R^2>0.98 across TPMS geometries. Generated scaffolds reproduced target features with similar accuracy (relative error <5%), while the covariance loss improved diversity, particularly for IWP scaffolds, increasing median coverage within the 5% tolerance region from 15.3% to 44.6%. Additional finite element simulations confirmed that generated designs generally matched target properties, with relative errors mostly <6%. This framework enables generation of diverse, mechanically and morphologically controlled TPMS candidates, supporting efficient inverse design for bone tissue engineering applications.