Physics‐Guided Deep Learning Agent for Pre‐Fabrication Recovery of Compromised Digital Geometries in Resilient Cyber Manufacturing

Evasive anomalies in digital geometries, such as stealthy cyber‐physical attacks on digital geometries and errors in geometry reconstruction from scanned data, can compromise functionality and cause unexpected failures in fabricated parts. Subtle changes to the internal surfaces of the complex geometry or creating small voids in the interior region are particularly difficult to detect either before or after fabrication. To enhance the resilience of cyber manufacturing systems against such anomalies in digital geometry data, this work proposes an efficient physics‐guided deep learning approach for pre‐fabrication detection of malicious geometric alterations and recovery of the original geometry. Specifically, we introduce a physics‐based geometric anomaly detection agent to localize potentially compromised regions in the spatial domain and propose a geometry recovery agent based on U‐Net deep learning model to reconstruct the original geometry within the identified local regions. The proposed geometry recovery framework is demonstrated and validated on both 2D and 3D geometry recovery testbeds, demonstrating its robustness and advancing toward in situ detection for true operational resilience against geometric cyber‐physical threats.