Machine learning applications in welding processes: Progresses and opportunities

Abstract

The increasing demand for intelligent and autonomous manufacturing has driven the integration of machine learning (ML) into modern welding processes. Owing to its ability to model nonlinear and cross-scale interactions and extract critical features from complex, high-dimensional data, ML is rapidly transforming the design, monitoring, and evaluation of welding processes. Based on this, the paper systematically reviews research progress in ML for four representative welding processes (arc, laser, resistance and friction stir welding) over the past decade. First, typical welding tasks are categorized into three domains: pre-weld design, in-process monitoring, and post-weld quality assessment. It then elaborates on the types of welding data used and their input-output relationships across different tasks and analyzes the architecture and algorithmic characteristics of mainstream ML models. Cross-process comparison reveals that the physical nature of each welding process determines the focus of ML research, model selection, and performance metrics. The study quantitatively compares the task-specific metrics of various models and presents successful industrial application cases. Despite significant progress, challenges persist in constructing high-quality and standardized datasets, improving model interpretability and generalization, and achieving robust real-time control in dynamic industrial environments. Based on the summarized emerging challenges, the perspectives on further development direction of applying ML in intelligent welding are also discussed.