Cross-Layer Edge Analytics and Physics-Informed Neural Networks for High-Frequency Defect Mitigation in Precision Metal Additive Manufacturing

• Ada John

  ladoke Akintola university of technology

  Author

High-frequency defects in laser powder bed fusion (LPBF) metal additive manufacturing continue to compromise part qualification and industrial adoption despite advances in in-situ monitoring. Existing machine learning approaches operate predominantly on single-scale data streams and lack physics-constrained inference capabilities at the edge, resulting in latency bottlenecks and limited generalization across defect classes. This study develops and validates a cross-layer framework integrating physics-informed neural network (PINN) surrogates with multi-tier edge-fog-cloud analytics for real-time defect detection and mitigation in LPBF. The methodology combines multisensor thermal, acoustic, and optical data streams processed through hierarchical PINN surrogates deployed on edge accelerators, with Bayesian calibration at the cloud layer. Experimental validation on NIST AM-Bench benchmark datasets for IN625 and Ti-6Al-4V demonstrates a macro-averaged F1-score of 0.984 across six defect categories (porosity, lack-of-fusion, cracking, balling, keyholing, delamination), with 11.3ms inference latency on edge hardware. The framework achieves a mean defect-detection rate of 98.7% and reduces scrap rates by 34.6% in simulated production scenarios. Feature importance analysis identifies melt pool thermal signatures and acoustic emission frequency shifts as dominant predictors. These findings establish a replicable architecture for physics-constrained edge intelligence that addresses the latency-accuracy trade-off in high-frequency defect mitigation, offering practical pathways for closed-loop process control in precision metal AM.

• pdf

Copyright (c) 2026 Ada John (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.