Geometry-informed multimodal variational autoencoder for real-time prediction of properties for Ti–6Al–4V fabricated using PBF-LB
A geometry-informed multimodal variational autoencoder linear hybrid model (GMVAE) was developed to use in situ processing signals along with geometry information from laser scanning patterns to predict the mechanical properties of Ti-6Al-4 V thin-walled tensile samples with varying thicknesses fabr...
Gespeichert in:
| Veröffentlicht in: | Journal of intelligent manufacturing |
|---|---|
| Hauptverfasser: | , , , |
| Format: | Journal Article |
| Sprache: | Englisch |
| Veröffentlicht: |
11.10.2025
|
| ISSN: | 0956-5515, 1572-8145 |
| Online-Zugang: | Volltext |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Zusammenfassung: | A geometry-informed multimodal variational autoencoder linear hybrid model (GMVAE) was developed to use in situ processing signals along with geometry information from laser scanning patterns to predict the mechanical properties of Ti-6Al-4 V thin-walled tensile samples with varying thicknesses fabricated using laser powder bed fusion (PBF-LB) additive manufacturing (AM). Feature representatives were extracted from in-process photodiode sensor data using pre-trained deep convolutional neural networks (DCNN) via deep transfer learning, and geometrical descriptors were extracted from laser scanning patterns. The feature representatives from photodiode sensor data were compressed in latent space using a trained variational autoencoder (VAE). These representatives were then concatenated with geometrical descriptors into multimodal features to train a regression model that predicts mechanical properties. Accuracies of 98.9% and 94.2% and an R 2 value of 0.97 were achieved in predicting ultimate tensile strength and elongation to fracture of samples with the proposed model. Benchmarking against DCNN and VAE models were conducted to demonstrate the advancement in prediction performance for samples with different thicknesses. The introduction of the VAE structure to balance multimodal in situ features between sensor signals and laser scanning patterns, as well as the designed dataset splitting with stepped learning paradigm, resulted in a reduction in computational cost compared to conventional deep transfer learning, indicating this framework has potential for real-time PBF-LB AM quality diagnosis and control. |
|---|---|
| ISSN: | 0956-5515 1572-8145 |
| DOI: | 10.1007/s10845-025-02693-3 |