Generative adversarial network–enabled microstructural mapping from surface profiles for laser powder bed fusion

Abstract

Laser powder bed fusion (LPBF) is the dominant metal additive manufacturing technique due to its advantages in near-net-shape production of complex parts with high resolution. However, conventional quality control of LPBF-fabricated parts, including microstructure characterisation, often relies on trial-and-error experiments. These methods can be time-consuming, resource-intensive, and potentially destructive to specimens. This study introduces an image-to-image translation Cycle-consistent Generative Adversarial Network (CycleGAN)-based framework for generating statistically equivalent microstructures of LPBF-fabricated samples directly from corresponding as-printed surface inputs. The results demonstrate that the framework can effectively generate crystallographic and morphological features across 22 different process parameters for LPBF-fabricated pure nickel. The distribution of microstructural descriptors, such as grain size, grain shape, and even grain boundary misorientation angles, shows no significant difference from that measured by experiments. The generated microstructural mapping using image inputs with CycleGAN outperforms those from other generation methods on both qualitative and quantitative evaluations. The developed framework is material-agnostic and can be fine-tuned for other LPBF materials via transfer learning, providing potential applications in in-situ process optimisation and microstructure design in LPBF manufacturing.

Keywords

• Microstructure
• surface morphology
• generative adversarial network
• laser powder bed fusion