Fig. 5. The workflow of multi-objective active learning.

a The task is to design scaffolds with a better mechanical response—fixed elastic modulus (E) and maximized yield strength (Y). b The 3D convolutional neural networks (3D-CNNs) for predicting E and Y. c The generative model for targeted scaffold generation. The encoder q_φ (z | x) with parameters φ took the scaffold porosity matrix as input, and the decoder p_θ (x | z) with parameters θ could act as a generator for proposing new scaffolds based on the learned latent z representation. d Multi-objective active learning loop (MALL) for high-performance scaffold discovery. First, the sampling algorithm sampled new data points from the latent z representation. Second, the decoder reconstructed the corresponding scaffolds so that the 3D-CNNs could infer their mechanical properties. Third, the most suitable candidates were selected based on the predicted E and Y. Finally, the strain‒stress curves of the selected scaffolds were calculated by the finite element method (FEM). New data were either fed back to the dataset or 3D-printed for further experiments.