3D pore space reconstruction using deep residual deconvolution networks

The observation and analyses of cores are the bases of various studies on oil and gas exploration and development. However, the real natural cores suffer from their own instability and constant weathering or erosion as well as experimental damages, leading to the changes of their physical and chemical characteristics. Digital cores can address the above issues by core digitalization and then reusing the original core images or data without damaging the real samples. The 3D reconstruction of pore space actually is an important step for the construction of 3D digital cores. There are two main ways for the reconstruction of pore space including physical experimental methods and numerical reconstruction methods. Physical experimental methods usually are quite time-consuming and expensive while numerical reconstruction methods are relatively inexpensive and more efficient. With the flourishing development of deep learning and its variants, the reconstruction of pore space possibly can benefit from the strong inherent ability of extracting characteristics from training images (TIs) hidden in deep learning. This paper proposes a reconstruction method using a deep residual deconvolution network (DRDN), considered as a variant of deep learning, in which the characteristics of TI are learned by using constant residual convolution in training and then the pore space is reconstructed by adding the previous residual convolution results to each deconvolution layer. Compared to some other typical numerical reconstruction methods, our method shows its efficiency and practicability in the reconstruction of 3D pore space.