UC Santa Barbara

Collecting 3D microstructural information of materials poses significant challenges, being a process that is time-consuming and expensive. While advancements in serial sectioning instrumentation have expedited the acquisition of 3D microstructure data, the process of obtaining crystallographic information through electron backscatter diffraction (EBSD) imaging continues to be a bottleneck, limiting the overall rate of data collection. In this research, we explore physics-informed computer vision methods to generate high-resolution 3D microstructure data. These methods are a primary step in solving the cost and time challenges. EBSD is a scanning electron microscope (SEM) imaging modality that maps crystal lattice orientation by analyzing diffraction patterns. EBSD maps are used to determine the microstructural properties such as texture, orientation gradients, phase distributions, and point-to-point orientation correlations, all of which are critical for understanding material performance. EBSD maps contain information about crystal 3d orientation in Euler space that follows crystallography symmetry properties. However, it is difficult to compute rotational distance and symmetry of 3d crystal orientation in Euler space. To solve these unique computational challenges associated with rotational distance and symmetry, we developed a physics-inspired 3D deep learning framework that uses rotational symmetry and quaternion orientation space as priors to generate high-resolution microstructure. The proposed quaternion residual block self-attention neural network (QRBSA) with physics-guided crystal symmetry loss is used to super-resolve high-resolution 3D microstructure from sparsely sectioned EBSD maps. We demonstrate, both qualitatively and quantitatively, that integrating the physics of microstructure into the deep learning architecture and loss function significantly reduces superresolution synthesis error compared to standard deep learning networks and loss functions.

Additionally, we propose a 3D Generative Adversarial Network (GAN) framework known as M(Material)-GAN, which can be used to learn the morphologies of 3D grains and synthesize realistic grains in microstructures. The moment invariances are used to quantitatively evaluate the generated grains and real grains. The creation of synthetic 3D grains represents a foundational step towards generating comprehensive synthetic 3D microstructures through deep learning techniques. The data and methods developed are available to the broader research community through the UCSB BisQue platform.