Fig. 2.

Framework to generate truss lattices with given stiffnesses. The inverse model takes the 21 independent elastic constants of ${ℂ}^{*}$ as input and first predicts a posterior distribution of possible lattice topologies, from which one (composite) topology is sampled and passed, jointly with ${ℂ}^{*}$, to a second NN, which predicts the corresponding geometrical parameters. The stiffness of the proposed lattice candidate is then reconstructed with the (independently trained) forward model and compared to the target stiffness. Predictions for ${ℂ}_{1111}$ of the forward and inverse models are compared to the actual stiffness (obtained via FE modeling and the forward model, respectively) and the corresponding R² deviations indicated (evaluated on a test set of 30,000 truss lattices). Further details of the NN architecture, training schemes, and accuracy are summarized in SI Appendix, section 3.