Modeling and Prediction of Laser Cladding Layer Morphology with Deep Learning

An ESFM network-based prediction model for cladding layers was developed using deep learning. A molten pool image dataset was constructed, followed by image preprocessing and augmentation. Using EfficientNetV2 as the baseline, the Fused-MBConv and MBConv blocks were optimized by integrating the Channel Shuffle mechanism from ShuffleNetV2, yielding the improved ESFM model through structural adjustment. The model was trained and validated on the molten pool dataset for classification prediction, with training time and accuracy compared quantitatively. Experimental results demonstrate that the improved ESFM model achieves fast convergence and 96% classification accuracy for cladding layer quality, confirming its effectiveness. This work provides a reliable approach for laser cladding quality prediction and supports further research toward stable deposition and process optimization.