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

Researchers have developed WeldMamba, a physics-guided AI model that predicts the future state of molten pools in laser wire-feed welding 500 milliseconds in advance by analyzing historical images and process parameters. This lookahead capability addresses the critical challenge of sensor-to-actuator delays in closed-loop welding control systems, achieving 74.63% mIoU accuracy on a 43-sequence dataset.

This research addresses a fundamental challenge in automated welding: the inherent delay between sensing the current weld state and executing corrective actions. WeldMamba solves this by predicting future weld-pool geometry rather than merely reacting to present conditions, enabling proactive control adjustments that can significantly improve weld quality and consistency.

The model's architecture combines multiple complementary approaches. It processes historical coaxial images through visual encoders, incorporates welding process parameters and electrical wire-state signals, and uses patch-level temporal state space modeling to capture spatiotemporal dynamics. The addition of physics-informed constraints—including signed-distance-function supervision and keyhole motion awareness—ensures predictions respect the underlying physics of the welding process rather than learning purely statistical patterns.

For industrial manufacturing, this advancement has meaningful implications. Precision welding is critical in aerospace, automotive, and heavy equipment sectors where weld defects create safety and reliability risks. A 500-millisecond lookahead window provides sufficient time for closed-loop controllers to adjust laser power, wire feed rate, or travel speed before problems manifest. The ablation studies demonstrating that temporal history and keyhole motion awareness drive performance suggest the model captures genuine physical phenomena rather than statistical artifacts.

The 74.63% mIoU metric indicates room for improvement before deployment in safety-critical applications, but the research direction is sound. Future work likely includes deployment on edge hardware, testing across different material combinations, and integration with industrial welding equipment. This represents incremental but important progress in computational manufacturing rather than a breakthrough.

• →WeldMamba predicts molten pool geometry 500ms ahead, enabling proactive control instead of reactive corrections in laser welding systems.
• →Physics-guided constraints and temporal state space modeling proved critical contributors to prediction accuracy in ablation studies.
• →The 74.63% mIoU performance demonstrates feasibility but suggests engineering refinement needed before industrial deployment.
• →The model processes multimodal inputs—images, process parameters, and electrical signals—to capture complete welding dynamics.
• →Addressing sensor-to-actuator delay through predictive modeling could improve weld quality and consistency across precision manufacturing applications.