Privacy-Preserving Porosity Prediction Using Spatiotemporal Graph Neural Networks in Laser-Based Additive Manufacturing Processes

Abstract

Porosity control in the LENS (Laser Engineered Net Shaping) additive manufacturing process is critical for ensuring structural integrity and durability, especially in high-performance applications. Traditional predictive models, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), often struggle to accurately capture the complex layer-by-layer connectivity in in-situ monitoring. This leads to decreased accuracy in porosity predictions and impacts quality certifications. Moreover, centralized machine learning approaches raise significant privacy concerns due to the proprietary nature of manufacturing data and the lack of robust mechanisms to protect sensitive information during model training. To address these challenges, this study proposes a unified Spatiotemporal Graph Neural Network (ST-GNN) framework that integrates spatial-temporal modeling and differential privacy (DP). The architecture combines Graph Convolutional Networks (GCNs) and Recurrent Neural Networks (RNNs) to capture spatial and temporal dependencies while employing differentially private stochastic gradient descent (DP-SGD) to protect data confidentiality during training. Experiments demonstrate superior accuracy in predicting porosity labels and sizes compared to baseline models. Additionally, privacy impact is quantified by measuring the privacy budget under different configurations, validating robust data protection with minimal performance trade-offs. This approach offers a privacy-preserving solution for quality certification, advancing secure, data-driven innovations in additive manufacturing.