Multi-task learning neural network (MB-Hybrid Net) for edge crack prediction and stress evolution in rolling of nickel-based super alloy ultra-thin strips

Traditional neural networks struggle to accurately model deformation behavior in nickel-based superalloy ultra-thin strip rolling. This study proposes a novel multi-task learning neural network (MB-Hybrid Net) with a multi-branch architecture to fuse heterogeneous process data, integrating LSTM, 1D CNN, and DNN components for simultaneous prediction of edge crack modes (none/left/right/bilateral) and eight stress distribution features. The model achieves 96.8 % classification accuracy and an R² of 0.967–0.994 for stress parameter prediction on the test set, outperforming conventional models by approximately 19 % in average R². SHAP-based interpretability analysis identifies the lateral hydraulic pressure distribution (T-hc-5/3/1) and positive bending force (F-bend-pos) as dominant factors governing edge crack formation, revealing direct mechanistic links between process parameters and edge quality. By synergizing data-driven learning with physical insights, the MB-Hybrid Net not only enables high-precision, dual-output prediction but also uncovers critical control variables for process optimization. This mechanism-informed hybrid modeling approach provides a robust foundation for intelligent rolling systems in ultra-thin strip manufacturing (thickness ≤ 0.2 mm).