Machine learning based hierarchy of causative variables for tool failure in friction stir welding. (15th June 2020)

• Record Type: Journal Article
• Authors: Du, Y.; Mukherjee, T.; Mitra, P.; DebRoy, T.
• Is Part Of: Acta materialia. Volume 192(2020)
• Page Start: 67
• Page End: 77
• Publication Date: 2020-06-15
• Subjects: Friction stir welding -- Aluminum alloys -- Tool failure -- Machine learning -- Neural network -- Decision tree; Materials -- Periodicals; Materials science -- Periodicals; Materials -- Mechanical properties -- Periodicals; Metallurgy -- Periodicals; Chemistry, Inorganic -- Periodicals; 620.112
• DOI: 10.1016/j.actamat.2020.03.047
• Languages: English
• ISSNs: 1359-6454

Abstract

Abstract: Since friction stir welding tools fail in service under various mechanisms, it is difficult to mitigate tool failure based on mechanistic understanding alone. Here we use multiple machine learning algorithms and a mechanistic model to identify the causative variables responsible for tool failure. We analyze one hundred and fourteen sets of experimental data for three commonly used alloys to evaluate the hierarchy of causative variables for tool failure. Three decision tree based algorithms are used to rank the hierarchy of the relative influence of six important friction stir welding variables on tool failure. The maximum shear stress is found to be the most important causative variable for tool failure. This is consistent with the effect of shear stress on the load experienced by the tool. The second most important factor is the flow stress which affects the plasticized material flow around the tool pin. All other variables are found to be significantly less important. Three algorithms also generate identical results and predict tool failure with the highest accuracy of 98%. A combination of mechanistic model, machine learning and experimental data can prevent tool failure accurately. Graphical abstract: Image, graphical abstract