Abstract Metal additive manufacturing is an efficient technique for producing metallic components with high flexibility. Understanding the mechanical properties of printed parts requires detailed analysis of their microstructures. While experiments are reliable for capturing as-printed microstructures, they are often costly and time-consuming. Alternatively, numerical simulations have become valuable tools to reduce the reliance on trial-and-error experimentation, powered by recent advances in high-performance computing such as GPUs for acceleration. However, a key question remains of effectively combining high-fidelity, yet expensive, experiment data with relatively lower-fidelity, yet massive, simulation data to achieve a coherent prediction of microstructures under various manufacturing process conditions. In this study, we leverage the power of deep generative models and propose a sim-to-real framework based on denoising diffusion probabilistic model (DDPM) to generate realistic microstructures for additively manufactured products. Specifically, we pre-train a DDPM on a large dataset of phase-field (PF) simulated microstructures. This model is then fine-tuned or distilled using a relatively small set of microstructure images obtained from electron backscatter diffraction (EBSD) experiments, with the goal of enhancing the authenticity of the generated images. The microstructures generated by our sim-to-real diffusion models show strong agreement with experimental results, evaluated using both machine learning and physics-based metrics. In particular, the sim-to-real diffusion models show accurate prediction under unseen experimental manufacturing process conditions (but covered by simulation data), demonstrating their excellent generalization ability.
Xie et al. (Tue,) studied this question.