摘要：AbstractThe nature of the measured data varies among different disciplines of geosciences. In rock engineering, features of data play a leading role in determining the feasible methods of its proper manipulation. The present study focuses on resolving one of the major deficiencies of conventional neural networks (NNs) in dealing with rock engineering data. Herein, since the samples are obtained from hundreds of meters below the surface with the utmost difficulty, the number of samples is always limited. Meanwhile, the experimental analysis of these samples may result in many repetitive values and 0s. However, conventional neural networks are incapable of making robust models in the presence of such data. On the other hand, these networks strongly depend on the initial weights and bias values for making reliable predictions. With this in mind, the current research introduces a novel kind of neural network processing framework for the geological that does not suffer from the limitations of the conventional NNs. The introduced single-data-based feature engineering network extracts all the information wrapped in every single data point without being affected by the other points. This method, being completely different from the conventional NNs, re-arranges all the basic elements of the neuron model into a new structure. Therefore, its mathematical calculations were performed from the very beginning. Moreover, the corresponding programming codes were developed in MATLAB and Python since they could not be found in any common programming software at the time being. This new kind of network was first evaluated through computer-based simulations of rock cracks in the 3DEC environment. After the model’s reliability was confirmed, it was adopted in two case studies for estimating respectively tensile strength and shear strength of real rock samples. These samples were coal core samples from the Southern Qinshui Basin of China, and gas hydrate-bearing sediment (GHBS) samples from the Nankai Trough of Japan. The coal samples used in the experiments underwent nuclear magnetic resonance (NMR) measurements, and Scanning Electron Microscopy (SEM) imaging to investigate their original micro and macro fractures. Once done with these experiments, measurement of the rock mechanical properties, including tensile strength, was performed using a rock mechanical test system. However, the shear strength of GHBS samples was acquired through triaxial and direct shear tests. According to the obtained result, the new network structure outperformed the conventional neural networks in both cases of simulation-based and case study estimations of the tensile and shear strength. Even though the proposed approach of the current study originally aimed at resolving the issue of having a limited dataset, its unique properties would also be applied to larger datasets from other subsurface measurements.Graphical abstractDisplay OmittedHighlights•Conventional NNs not extract the information in the data to its full capacity.•New kind of network structure introduced by rearranging all the basic elements.•Challenges of conventional NNs addressed in dealing with limited/repetitive data.•Machine learning Applied in investigating tensile and shear strength of the rock.•New network structure validated through simulation and a case study.
关键词：KeywordsenTensile strengthShear strengthGas HydrateFeature engineeringRock engineering dataNeuron model
