Improving Friction Stir Welding Reliability Through Machine Learning Control and Physics-Based Modeling

Self-reacting friction stir welding (SR-FSW) is a solid-state joining technique utilized in the construction of the SLS. Compared to fusion welding, SR-FSW offers the potential to maintain the high strength of heat treatable aluminum alloys by reducing the temperature rise that occurs near the weld. However, challenges related to SR-FSW have persisted and abnormally low mechanical tests have been sporadically observed. An extensive joint investigation by MSFC's Welding and Manufacturing Team and the NESC has identified a strong relationship between input power and weld quality. This insight suggests that weld reliability could be improved by ensuring input power is maintained within a target window. Here we explore two technical approaches achieve that goal: using physics-based models to inform better process design and using novel control strategies to perform in-situ process optimization.

The physics-based model uses computational fluid dynamic (CFD) principles to predict material flow and plastic heating and has been shown to predict input power and wormhole defects.

The power control capability uses artificial intelligence and machine learning (AI/ML) to adjust the operational process parameters in real-time. A reinforcement-learning based agent continuously and autonomously learns the weld response to process changes under dynamic operating conditions. Initial results of AI-controlled welds show strong promise with clear learning and improvement after a single weld.

Together, these digital transformation developments can create a more robust process to weld flight hardware. Ultimately, optimized process design and control will minimize the need for repairs and rewelds and reduce production costs and time.