GNNs for mechanical properties prediction of strut-based lattice structures

The mechanical properties of strut-based lattice structures are greatly influenced by cell topology, which can be modified by changing connections between nodes within a single unit cell. However, since cell topology is not a continuous variable and varies in non-Euclidean space, it is difficult to provide a quantitative relationship between cell topology and mechanical properties. Here, we presented a graph-based deep learning approach for mechanical property prediction of lattice structures with a message passing neural network (MPNN). A dataset of over 100,000 cell topologies was first generated using a proposed exhaustive algorithm. The MPNN model was trained and tested using simulated compressive strength data of lattice panels with 1980 cell topologies randomly selected from the topology dataset. The mean absolute percentage error on the test dataset reached 8.82 %. Based on the trained MPNN model, 10 cell topologies corresponding to the highest predicted compressive strength at different relative densities were used to manufacture test specimens by powder bed fusion technique. The test specimens exhibited higher compressive strength than most typical lattices. This work reveals a potential for applying graph-based deep learning techniques on property prediction and topology optimization of strut-based lattice structures.