A Deep Learning-based Data-driven Approach for Modeling and Optimization of Laser Transmission Welding of Polypropylene

Abstract

In this study, a novel multi-stage framework is explored for laser transmission welding of polypropylene by integrating the design of experiments (DoE), artificial neural networks (ANN), non-dominated sorting genetic algorithm-II (NSGA-II), and multi-objective optimization by ratio analysis (MOORA). The framework enables comprehensive experimental investigation, process modeling, and multi-objective optimization. The response surface method (RSM) based DoE is used to develop correlations between welding parameters and responses, which form the foundation for experimental investigations. ANN models, incorporating additional fractional factorial DoE data, are employed for precise non-linear mapping of process parameters and responses, with predictive accuracy surpassing that of RSM models. The 3-6-1 ANN architecture is demonstrated to predict weld strength with high precision, while the 3-7-2-1 model is found to predict weld width accurately. These ANN models are used as objective functions for simultaneous optimization via NSGA-II, generating Pareto-optimal sets. These sets are further prioritized by MOORA, with an optimal parameter set of 220 W laser power, 81.29 mm/s scanning speed, and 63.97 mm defocus distance, yielding a weld strength of 63.86 N/mm and a weld width of 3.24 mm. The proposed synergistic DoE-ANN-NSGA-II-MOORA framework not only confirms its efficacy in this particular case but is also adaptable for other materials and processing applications.