Artemis Friction Stir Weld Process Optimization Using Machine Learning and Informatics

The Artemis core stage self-reacting friction stir welding (SRFSW) process was optimized to significantly reduce process variation and improve weld strength. Machine learning models were used to establish process-structure-property relationships and identify an optimal process window. A space-filling experimental design enabled exploration of the design space while capturing nonlinear effects.

To support this effort, a data pipeline was developed to automatically ingest, verify, and clean diverse data sources, including fracture surface images, sequential weld tool sensor data, and tabular processing and property data from test labs. Fracture surface defects were quantified using a convolutional neural network (CNN) for segmentation, coupled with traditional computer vision for feature extraction. Sequential sensor data was analyzed using long short-term memory (LSTM) neural networks, with activation maps providing insight into model interpretability.

Ultimately, an ideal power input window was identified, leading to improved weld quality. Data-driven and physics-based models linking process parameters to power input were used to establish a controllable process window to achieve optimal welds.

This presentation will explore how the integration of machine learning, experimental design, and informatics enabled data-driven hypothesis generation and validation, leading to measurable improvements in weld quality.