Multitask 3D Convolutional Neural Network–Based Multiphysics Design Exploration of Triply Periodic Minimal Surface Bone Scaffolds

Triply periodic minimal surface (TPMS) scaffolds have emerged as promising candidates for bone tissue engineering due to their ability to balance mechanical stiffness, mass transport, and flow-induced shear environments. However, accurate evaluation of these coupled properties typically relies on computationally intensive finite element and computational fluid dynamics simulations, which limits large-scale exploration of the TPMS design space. In this study, we propose a multitask three-dimensional convolutional neural network (3D-CNN) surrogate model for the concurrent prediction of key effective properties of TPMS bone scaffolds, including apparent elastic modulus, intrinsic permeability, effective diffusivity, and a wall shear stress (WSS)–based exposure metric, directly from voxelized scaffold geometries. A dataset of 1,000 unique TPMS designs spanning gyroid, Schwarz-P, and diamond families was generated and labeled using multiphysics simulations, with additional instances obtained through symmetry-preserving transformations. The surrogate model employs a physically informed seven-channel geometric representation and joint learning across tasks to capture shared structure–property relationships. On a held-out test set, the multitask surrogate demonstrated robust predictive performance across all targets, achieving coefficients of determination up to 0.53-0.55 for transport and mechanical properties in challenging intermediate-porosity regimes, while reducing root mean squared error by approximately 40–55% compared to classical analytical models. Relative to single-task CNNs, multitask learning further reduced prediction errors by 7-12%, particularly for transport- and shear-related quantities. Family-wise and representative-design analyses revealed physically consistent topology-dependent trends, with Schwarz-P structures favoring stiffness, gyroid architectures promoting transport performance, and diamond scaffolds offering balanced trade-offs. Pareto-based exploration confirmed that no single topology is universally optimal, underscoring the need for application-specific scaffold selection.