Spatiotemporal Graph Neural Networks For Layer-Wise Defect Propagation Analysis

Monday, October 20, 2025: 3:20 PM
Ms. Temilola Gbadamosi-Adeniyi , North Carolina State university, Raleigh, NC
This research introduces a novel framework employing Spatiotemporal Graph Neural Networks (ST-GNN) to model and predict defect propagation across multiple build layers in electron beam powder bed fusion (EB-PBF) processes in real time. Traditional monitoring approaches have limited capacity to anticipate how defects evolve through subsequent layers, creating a critical gap in quality assurance for additive manufacturing.

Our methodology transforms electron emission data into a graph-based representation where nodes encode feature vectors from the emission data, process parameters, and local geometry, while edges establish spatiotemporal relationships between regions. The core architectural innovation lies in our specialized ST-GNN implementation, which integrates spatial convolution operations with temporal attention mechanisms and message passing layers to capture complex defect evolution dynamics.

The framework incorporates physics-informed training protocols with progressive learning strategies and multi-task optimization techniques. Our custom loss functions embed domain knowledge of thermal physics, enabling the model to learn the underlying physical phenomena governing defect propagation.

Validation experiments involving controlled defect introduction, cross-sectional analysis, and micro-CT scanning demonstrate the framework's capability to predict defect evolution with high accuracy while maintaining temporal consistency. Performance evaluation across multiple metrics confirms the approach's effectiveness within real-time processing constraints.

This research addresses a fundamental challenge in additive manufacturing quality control by providing predictive capabilities that enable proactive intervention before defects propagate through subsequent layers. The framework's ability to integrate physical domain knowledge with data-driven learning presents a significant advancement for in-situ monitoring systems in additive manufacturing processes.