A high-fidelity digital rock representation based on digital grinding combined with deep learning for four-dimensional lattice spring model

This paper introduces a method for constructing high-fidelity digital rock using digital grinding and deep learning , specifically for the Four-Dimensional Lattice Spring Model (4D-LSM). Initially, rock sequence images are captured with a self-designed digital grinding equipment. Bicubic interpolation is then applied to fill missing pixels, ensuring uniform resolution. The images are subsequently deblurred using DeblurGAN, a deep learning network trained with existing high-definition images. This process results in high-fidelity 3D true-color digital rock geometry reconstruction. An Artificial Neural Network (ANN) identifies mineral components, which are then mapped into the 4D-LSM to create the high-fidelity 3D true-color Grain-Based Model (GBM). The mechanical behavior of the GBM is analyzed using the 4D-LSM, incorporating strength reduction factors which can be easily calibrated through a modified Newton algorithm. Results demonstrate that the high-fidelity 3D true-color GBM accurately replicates the mechanical behavior and failure processes of granite, offering improved consistency with experimental data compared to homogeneous models . • Utilization of deep learning, specifically DeblurGAN, enhances the high-fidelity of digital rock images by rectifying pixel loss and resolution issues inherent in digital grinding. • Integration of Artificial Neural Networks (ANN) with digital rock technology enables the precise mapping of mineral components to develop a detailed 3D true-color Grain-Based Model (GBM). • The application of a modified Newton iteration algorithm ensures precise calibration of strength reduction factors, automating the parameter selection for the 3D true-color GBM 4D-LSM.