| Abstract Scope |
Introduction
The use of lightweight material combinations has been highly demanded in manufacturing automotive structures. However, making robust dissimilar material joints of such lightweight materials is still challenging. A significant barrier to achieving high-quality and repeatable joint performance is a deficient understanding of the relationship between the welding process, joint attributes, and joint performance. The present study proposes and demonstrates a novel machine learning (ML) method to establish process-structure-property relationship in resistance spot welding, meanwhile addressing two major issues with the existing methods for weld quality prediction. First, in contrast with a single point prediction of joint quality, the proposed ML method is capable to predict not only the statistic average of joint performance properties, but also its scattering (quality variation) with high level confidence. Second, most existing methods for weld quality prediction are confined to a given material combination or certain process conditions. Our proposed method utilized one unified ML framework to analyze a variety of input parameters and output parameters related to RSWs. The model has been found to progressively gain comprehensive learning as it expands to cover more weld stack-ups, base materials, welding conditions, etc. Such unified and expandable ML architecture makes it possible to guide RSW process development with ‘untested’ materials, thickness, and other conditions.
Approach
The unified model design strived to (i) improve the wisdom of welding engineering by designing data representation with support of welding physics knowledge and to (ii) use the ML analysis results to provide insights for resistance spot welding of Al and steel alloys. The supervised deep neural networks (DNN) were developed to establish the quantitative correlation between weld process conditions, joint attributes, and joint performance. The data streams being analyzed comprised of a wide range of welding data associated with Al and steel resistance spot welded joints, including material specifications, welding process parameters such as current, clamp load, and duration, etc.; microstructural features such as hardness, interface topology, and intermetallic compound (IMC), etc.; and bulk performance in terms of peak load, elongation, and energy under different loading/mechanical testing conditions (e.g., coach peel, lap shear, cross tension tests). A novel Machine Learning method has been proposed to predict not only joint performance but also weld repeatability based on a fairly large dataset of dissimilar Al-steel resistance spot welds collected in General Motors research lab over several years of research and testing. The data set consisted of over 20 different material combinations and more than 800 welding conditions. The model utilized one neural network design and one training strategy for all material combinations and weld stack-ups.
Results and discussion
To illustrate the effectiveness of the DNN model, a regression analysis has been performed between the measured and ML predicted peak load, extension at break, and total energy for validation testing welds only (a total of 40 weld stackups). High correlation coefficients (greater than 0.9) between the measured and predicted values were obtained, indicating the DNN model identified the high dimensional correlations among the weld process conditions/joint attributes and mechanical performance of RSW joints. All cases of different weld stackups show promising and reasonable prediction results.
The performance of the unified ML modeling framework was further tested with an additional dataset of “unseen or independent” welds. These independent welds were from material/stackup combinations that were never used in the development of ML model. The ML model predicted the joint performance with greater than 80% accuracy, for over 90% independent testing welds. Generally, the unified ML framework exhibited its strong predictive and generalization capability when exposed to unseen weld stackup.
The unified ML model was then applied to determine the weld process conditions/attributes which would meet a specific target performance (high performance with good repeatability) for a certain weld stackup, as well as to search for possible optimal performance. This was done by utilizing an optimization scheme in connection with the fully-trained ML framework. In addition, sensitivities of the joint quality and its repeatability with respect to the key weld variables were also studied.
Conclusion
The present work proposed a unified ML model/architecture that is expendable to a wide range of material combinations and other conditions. Its effectiveness has been demonstrated in modeling complex relationships between RSW process parameters, weld attributes, and joint performance. The ML model predicted both average joint performance and its scattering with high level confidence. The unified ML framework shows great potential for practical application for developing and optimizing new joining process/substrate combinations to achieve defect-free, high performance, and repeatable joints in a more cost-effective and accelerated way. Even though a considerable amount of analyses were carried out for validation, more work is still required for methodology verification, especially for ‘untested’ materials. |