Machine learning predictions on the compressive stress–strain response of lattice-based metamaterials

Predicting the stress–strain curve of lattice-based metamaterials is crucial for their design and application. However, the complex nonlinear relationship between the mesoscopic structure of lattice materials and their macroscopic mechanical behavior makes prediction challenging. In this study, beam element models of over 20,000 lattice structures were established using Python scripts, and calculations were performed in ABAQUS to obtain training and testing datasets. The spatial features of each lattice-based metamaterial were then encoded into a graph, a data structure recognizable by machine learning algorithm. Utilizing machine learning methods, a Structure to Sequence Neural Network was constructed and trained, achieving rapid prediction of the compressive stress–strain curves for lattice-based metamaterials. Afterwards, several lattice structures were randomly selected and 3D printed. The accuracy of the simulation results as well as machine learning predictions was validated through quasi-static compression experiments. It is revealed that the proposed Neural Network model outperforms the traditional Artificial Neural Networks as the errors are reduced while the Coefficient of Determination is higher. The results demonstrate the accurate fitting between the complex spatial features of the lattice-based metamaterials and their stress–strain curves, which provides a potential methodology for inverse optimization of the lattice-based metamaterials in the future.