Scaling Deep Learning for Material Imaging: A Pseudo-3d Model for Tera-Scale 3d Domain Transfer

The characterisation of the internal 3-dimensional (3D) structure of complex materials has been revolutionised with the use of deep-learning artificial intelligence. However, this process often requires neural networks that must be trained every time and remains computationally limited to 2D or relatively small 3D domains (10^8 voxels). Herein, we introduce a Pseudo-3D domain transfer network (P3T-Net) for transfer capable of inference on domains approaching the tera-scale (10^12 voxels), solving domain mismatch between trained networks and other inference inputs, third-axis misalignment of 2D networks on 3D data, and the limited inference size of memory-inefficient 3D networks on large-scale 3D data. Solving these interconnected challenges has enabled to use deep learning for multi-modal 3D large-scale nano/micro-CT imaging. Significant applications of domain transfer are demonstrated with both pixelwise accuracy and validation of physical metrics. These include (i) image quality enhancement of fast-scan for geological rock and hydrogen fuel cells, (ii) transferring single-mode images to dual-mode quality for lithium-ion battery which reduces the imaging time by 60%, (iii) accurate multi-label segmentation of images under different imaging conditions, and (iv) efficient large-scale 3D inference (10^12 voxels) on a single GPU. The outcomes of each application highlight the ability of P3T-Net to accurately and efficiently transfer 3D cross-domain images for a variety of materials in diverse conditions.