Performance Evaluation of TWIST Welding Using Machine Learning Assisted Evolutionary Algorithms

The Transmission welding using incremental scanning technique (TWIST) combines linear feed with an oscillating laser beam to enhance weld quality and expand the process window. However, TWIST welding is influenced by nonlinear process variables, and achieving multiple objectives concurrently is challenging due to conflicting performance attributes. In industrial practice, time constraints and project specifications limit the effectiveness of methodologies tailored to specific workpiece materials or single performance optimization. The present study employs an artificial neural network (ANN) to establish a correlation between TWIST welding parameters and desired performance attributes. Various ANN model architectures are evaluated, with the 5-11-6-2 architecture achieving the highest accuracy (correlation coefficient of 0.998). For multi-objective optimization, the non-dominated sorted genetic algorithm (NSGA-II) and non-dominated sorted teaching learning-based optimization (NSTLBO) algorithm are employed, utilizing the ANN model's fitness function as the objective. The newly developed two-step model provides operators with the flexibility to prioritize factors based on project requirements, resulting in improved outcomes. Comparative analysis of the algorithms using seven metrics demonstrates that NSGA-II outperforms NSTLBO in solution prediction, albeit with slightly increased computing time. NSGA-II offers a broader range of Pareto optimum solutions compared to NSTLBO, which converges narrowly and restricts non-dominated sets. Validation experiments confirm the adequacy of both algorithms, supporting the effectiveness of the two-step model. The proposed methodology enables practitioners to achieve better weld quality, accommodate conflicting performance attributes, and effectively optimize multiple objectives in industrial applications.