Prediction of laser welding molten pool light intensity distribution based on an artificial neural network

This study aims to develop a spatio-temporal predictive model for the luminous intensity distribution of the laser welding molten pool-a key visual indicator of process stability and quality-to overcome the limitations of conventional analytical models in handling complex multi-physical interactions. A data-driven framework based on a nonparametric artificial neural network architecture is proposed. Gaussian functions are employed as radial basis functions to capture localized spatio-temporal variations in the light field. The root mean square error is adopted as the evaluation metric and integrated into a systematic hyperparameter optimization procedure to enhance model fidelity and robustness. The optimized model successfully predicts two distinct molten pool luminous patterns under different welding conditions. Predictions show strong agreement with synchronized high-speed experimental images, confirming the model's accuracy and generalization capability. This method effectively reconstructs the molten pool's luminous signature, demonstrating significant potential for real-time process monitoring, online anomaly detection, and non-destructive quality assessment in advanced laser welding operations.