Mechanism Informed Image Feature Decoding for Melt Pool Morphology Evolution Prediction in Laser Powder Bed Fusion

Accurate prediction of the melt pool morphology evolution is a crucial foundation for realizing control of melt pool and reducing laser powder bed fusion process defects. The existing methods do not effectively consider both the evolution mechanism and the influence of the historical scanning trajectory. A novel mechanism informed image feature decoding approach is proposed, which achieves prediction by guiding an encoder-decoder framework with melting mechanisms. The material properties, process parameters, and scanning strategies are integrated by a mechanism analytical model to obtain regional theoretical temperatures, and convolutional neural network model is used to extract temperature distribution features. The gated recurrent unit is used to fuse the mechanism feature, laser moving direction feature, and sensing image features and describe their evolution along the scanning track. Moreover, the relationship between the melt pool evolution and the historical scanning track is modeled as a function of time difference and spatial distances by an improved attention mechanism. The proposed framework is validated by predicting the melt pool morphology evolution under varying geometry and laser power. The experimental results indicate that the method can better predict the melt pool morphology evolution. Compared to sensing image features and mechanism feature, the mean absolute percentage error decreased by 3.5% and 1.8%, respectively. The analysis results based on the Shapley additive explanations further indicate that the prediction model has good mechanism consistency.