.png){kind=icon}
.png){kind=icon}
Return
Hybrid PINN–Bayesian multi-fidelity learning for cross alloy defect severity modeling and scalable control in metal additive manufacturing
.png){kind=icon}
PRE
AI
N
DOI：10.1007/s40964-026-01896-1.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Laser Powder Bed Fusion (LPBF) additive manufacturing enables high-precision fabrication of complex metal components; however, defect formation driven by unstable melt-pool dynamics remains a critical challenge. Accurate prediction of defect severity is essential for improving process reliability and material performance. In this study, a physics-aware hybrid framework named HyPhAIM is proposed for predicting the Defect Severity Index (DSI) in LPBF systems. The framework integrates physics-informed neural networks (PINNs), Bayesian optimization, and multi-fidelity learning to enhance predictive accuracy, physical consistency, and generalization capability under limited data conditions. Using the Melt Pool Defect Analysis Dataset (Kaggle), the proposed model learns complex relationships between process parameters, thermal behavior, and defect formation mechanisms while embedding thermodynamic constraints to ensure physically meaningful predictions. The model achieved strong predictive performance for DSI estimation, with R2 = 0.987, RMSE = 0.42, MAE = 0.33, MAPE = 2.7%, and Pearson correlation = 0.993, demonstrating high accuracy and robustness. Furthermore, the framework maintains computational efficiency with a total runtime of 9.6 s. To evaluate generalization, external validation using the NIST AM-Bench dataset confirms the model’s ability to adapt to unseen LPBF conditions with minimal performance degradation. The results highlight that integrating physics-informed learning with uncertainty-aware optimization significantly improves defect prediction reliability. Overall, HyPhAIM provides a scalable and transferable approach for physics-consistent defect severity prediction, offering a promising direction for intelligent and data-efficient additive manufacturing systems.
Keywords:
Additive manufacturing
Physics-informed neural network
Bayesian optimization
Multi-fidelity learning
Defect prediction
Journal
P
IF:
5.4
Papers:
1.8K
Citations:
3.2K
Organization
Cited Papers