AI-based real-time prediction of cross-sectional phase distribution during laser heat treatment via sequential thermal imaging and image-to-image synthesis

Abstract

Real-time monitoring of laser heat treatment is essential for ensuring microstructural consistency and process stability, but remains challenging due to complex thermal behavior and phase transformations. While deep learning has been applied to various laser-based processes, no prior study has addressed the real-time prediction of cross-sectional phase distributions in laser heat treatment. This study proposes the first AI-driven framework for real-time phase mapping during laser heat treatment of S45C steel, using sequential infrared surface temperature images. A convolutional gated recurrent unit architecture was developed to extract both spatial and temporal features, enabling pixel-wise image-to-image translation from thermal images to phase maps. Experimental data were collected from 16 experimental conditions by varying laser power and scanning speed. The model was trained using cross-sectional phase maps obtained from metallographic analysis, achieving high prediction accuracy across heat-affected, heat-treated, and melted regions. The model successfully captured key phenomena such as heat accumulation, geometric growth, and asymmetric phase evolution due to directional thermal gradients. This approach demonstrates a non-invasive, data-driven method for high-resolution monitoring of phase evolution in laser heat treatment. It offers new insights into thermally driven phase evolution and holds promise for integration into intelligent thermal processing and materials design strategies.