Introduction: Friction stir welds often suffer from premature tool failure and defects such as voids. Voids in the welded components affect the mechanical properties and serviceability of the joints. Both tool failure and void formation in friction stir welding (FSW) depend on multiple complex physical processes, and cannot be easily understood either by theory or experiments alone. Therefore, currently, there is no simple method to prevent premature tool failure and minimize the void formation in FSW based on scientific principles. A recourse is to use data-driven machine learning. In this research, we develop, test and utilize supervised machine learning algorithms to forecast the tool failure and void formation. Two machine learning algorithms, neural network and decision tree are used to identify conditions to prevent tool failure and void formation in FSW of aluminum alloys.
Technical approach: We collect one hundred and fourteen sets and one hundred and eight sets of independent experimental data from the literature on tool failure and void formation, respectively. These data are for three commonly used aluminum alloys, AA 2024, AA 6061 and AA 7075. The data are used to train, validate and test both the neural network and decision tree. First, the alloy properties and welding parameters such as welding speed, tool rotational speed, axial pressure, shoulder diameter, and plate thickness are used to forecast tool failure and void formation using neural networks. Then the hierarchical influences of several complex FSW variables on both tool failure and void formation are predicted using decision trees. These complex variables include temperature, strain rate, flow stress, traverse force, torque and maximum shear stress on tool pin and are calculated using a well-tested heat transfer and material flow model of FSW.
Results and discussions: Maximum shear stress on tool pin and flow stress are the most important variables that control tool failure. Similarly, void formation is mostly influenced by the peak temperature and maximum shear stress. The decision tree could predict tool failure with 98.1% accuracy, which is much better than the 87.0% accuracy obtained using the neural network. Predictions of both neural network and decision tree using the computed variables as inputs are more accurate than that using the raw welding variables.
Summary and conclusion: In summary, experimental data sets, model results, and two machine learning algorithms are used to predict appropriate conditions to prevent premature tool failure and voids formation. The complex FSW variables, temperature, strain rate, flow stress, torque, traverse force and maximum shear stress on tool pin, are superior to the raw welding parameters in predicting tool failure and void formation.