Microstructural Generative Inverse Models for Additive Manufacturing Digital Twins

Abstract

Metal additive manufacturing (AM) offers substantial advantages to the fabrication of complex structural components, but its widespread adoption is slowed due to complexities in the process-structure-property relations which is the cause for uncertainty and drives risk. Accurately quantifying these complex relations solely via experimental workflows is time consuming and costly. Integrated computational materials and engineering (ICME) frameworks address this challenge by leveraging advanced computational tools, such as crystal plasticity finite element models (CPFEM), to predict critical mechanical properties directly from microstructural data. However, these frameworks require a large datasets that are often difficult to obtain, and their predictive accuracy depends on the availability of realistic AM microstructure representations. Generative deep learning models are a promising approach towards synthesizing AM microstructures which contain complex morphological and crystallographic features. Denoising diffusion probabilistic models (DDPMs), in particular, show a strong potential to capture the inherent complexities of AM structures. In this work, we demonstrate that DDPMs can reliably synthesize digital-twin orientation-imaging micrographs, and we support this claim with quantitative evidence of the fidelity of the generated samples. Additionally, we establish a classifier-free guidance framework for solving complex inverse problems such as property-to-structure and process-to-structure mappings, and we showcase a potential application on a qualification-relevant uncertainty quantification problem. Our work highlights the potential of AI-based generative modeling as a key component of computational frameworks for accelerating AM materials qualification and certification.