Machine learning-based mechanical performance prediction and design of lattice structures

•A data-driven framework enables elastic modulus prediction and inverse design of lattice cells.•The designed neural network model achieves superior prediction accuracy.•The proposed C-VAE captures intrinsic correlations between lattice topology and mechanical properties.•An inverse strategy with im...

Ausführliche Beschreibung

Gespeichert in:

Bibliographische Detailangaben
Veröffentlicht in:	International journal of mechanical sciences Jg. 294; S. 110230
Hauptverfasser:	Liu, Yifan, Huang, Wei, Wang, Zhiyong, Zhang, Jie, Liu, Jiayi
Format:	Journal Article
Sprache:	Englisch
Veröffentlicht:	Elsevier Ltd 15.05.2025
Schlagworte:	3D-printed architected materials Conditional variational autoencoder Elastic modulus Inverse design Lattice unit cell Machine learning Conditional variational autoencoder Elastic modulus 3D-printed architected materials Inverse design Lattice unit cell Machine learning
ISSN:	0020-7403
Online-Zugang:	Volltext
Tags:	Tag hinzufügen Keine Tags, Fügen Sie den ersten Tag hinzu!

Beschreibung
Zusammenfassung:	•A data-driven framework enables elastic modulus prediction and inverse design of lattice cells.•The designed neural network model achieves superior prediction accuracy.•The proposed C-VAE captures intrinsic correlations between lattice topology and mechanical properties.•An inverse strategy with improved efficiency compared to traditional topology optimization has been developed. A data-driven framework is proposed that integrates automated geometric modeling and high-throughput finite element simulations to enable both forward prediction and inverse design of lattice unit cells. By establishing an efficient workflow, a substantial numerical simulation database is generated encompassing diverse truss-based topologies, the fidelity of which is corroborated by 3D-printed prototypes and quasi-static compression tests. A specially designed network is trained to predict elastic modulus, demonstrating improved accuracy (MSE = 0.046, R > 0.994) and reduced overfitting compared to graph neural network (GNN). Building upon this predictive model, a conditional variational autoencoder (C-VAE) is introduced that learns a low-dimensional latent space simultaneously conditioned on topological features and mechanical performance. Subsequent dimensionality reduction and clustering analyses, utilizing principal component analysis (PCA) and K-means algorithms, elucidated intrinsic correlations between rod connectivity and structural stiffness. Ultimately, by coupling the C-VAE with the predictive model, a high-throughput inverse design strategy is realized, enabling the fabrication of unit cells that achieve the prescribed elastic modulus with remarkable fidelity (design error < 2 %), and with a computational design time on the order of 10 s. And the design speed is approximately 66 times faster than that of traditional topology optimization methods. This paradigm significantly accelerates the design of 3D-printed architected materials, offering a promising and data-efficient approach for exploring complex structural geometries. [Display omitted]
ISSN:	0020-7403
DOI:	10.1016/j.ijmecsci.2025.110230