Intelligent generative design: a transfer learning-enhanced bidirectional mapping framework for customised parametric structural design

Tissue engineering scaffolds with tailored mechanical properties are crucial for regenerative medicine. However, establishing the intricate relationships between scaffold design parameters and resulting mechanical behaviour remains a challenge. Here, we present a transferable, machine-learning-driven framework for the rapid customisation of triply periodic minimal surface (TPMS) scaffolds. We employ a genetic algorithm to optimise a back-propagation neural network (GA-BPNN) for forward prediction of elastic modulus from structural parameters (porosity, pore size). Transfer learning enables efficient adaptation of the model to different TPMS topologies (Primitive, Gyroid, Diamond, I-Wrapped Package) with minimal retraining. Furthermore, we develop a BPNN-SCLFPSO (synchronous changing learning factor particle swarm optimisation) reverse search model, allowing for the inverse design of TPMS structures with targeted mechanical properties. Simulations demonstrate high predictive accuracy (R2 > 0.97) across diverse material compositions. Physical validation via additive manufacturing and compression testing confirms the model's reliability, with average relative errors of 7.6% and 4.8% for PCU and Ti alloy, respectively. This framework offers a powerful tool for on-demand design of TPMS scaffolds, accelerating the development of personalised tissue engineering solutions.