Deep learning based reconstruction of transient 3D melt pool geometries in laser powder bed fusion from coaxial melt pool images

Melt pool monitoring can provide guidance for quality control in additive manufacturing. Existing techniques focus on 2D melt pool characteristics. In this work, we explore in situ prediction and control of 3D melt pool geometries within a simulation environment. To reconstruct the transient 3D melt pool geometry with only the 2D coaxial image as an input, we adopt a U-net model trained with a synthetic image dataset generated from simulations. The results show that the root-mean-square error of the pixel-wise predicted melt depth is 1.14μm, demonstrating the potential of the model in enabling precise laser path-wise 3D melt pool control.