Flow-related defects in friction stir welding are critical for the joints affecting their mechanical properties and functionality. One way to identify them, avoiding long and sometimes expensive destructive and nondestructive testing, is using machine learning tools with monitored physical quantities as input data. In this work, artificial neural network and decision tree models are trained, validated, and tested on a large dataset consisting of forces, torque, and temperature in the stirred zone measured when friction stir welding three aluminum alloys such as 5083-H111, 6082-T6, and 7075-T6. The built models successfully classified welds between sound and defective with accuracies over 95%, proving their usefulness in identifying defects on new datasets. Independently from the models, the temperature in the stirred zone is found to be the most influential parameter for the assessment of friction stir weld quality.