Physics-Guided Spatiotemporal State Space Modeling for Lookahead Molten Pool Segmentation in Laser Wire-Feed Welding

The paper introduces WeldMamba, a physics-guided spatiotemporal state space network designed for real-time segmentation of weld pools in laser wire-feed welding. This model leverages historical grayscale images, welding parameters, and electrical signals to predict the future layout of key regions, achieving a mean Intersection over Union (mIoU) of 74.63% with a 500 ms lookahead on a 43-sequence dataset. The architecture incorporates advanced features such as patch-level temporal modeling and motion-aware decoding, which are crucial for improving closed-loop control in welding processes.