Flysch materials are a common source of slope instabilities and other geotechnical problems. Some attempts of characterizing such materials have been done, probing to be a challenging aspect. One European region where flysch materials are abundant is the Spanish Basque Arc Alpine region. A broad geological-geotechnical investigation was conducted on 33 locations spread in an area of approximately 100 km² in the Spanish Basque Arc Alpine region to geomechanically characterized the “Upper Cretaceous Flysch” materials found in the area. Such characterization was done following a GSI vs. uniaxial compressive strength of the intact rock chart given by other authors. Even though that procedure showed to be appropriate, values of Hoek-Brown parameters proposed based on such chart, showed not to match very well with the ones obtained in the laboratory for the 33 points analyzed. This caused the estimation of shear strength parameters of flysch materials to be poor. To improve it, Artificial Neural Networks were used. These artificial intelligence algorithms are of common use in engineering and enable to find no-linear correlations which may be difficult to find otherwise. Results show a very good performance when using Artificial Neural Networks, achieving determination coefficients R² between laboratory values and numerical ones close to 1.

The continuous convergence monitoring during and after excavation is an important tool in the application of the observational method for tunnels design (Schubert 2008. Geomech. Tunn. 1(5):352–357). Previous works have shown that, the direct analysis of convergence measurements with empirical models allows reliable long-term predictions of ground deformations. It is shown that, by monitoring convergence for a few tens of days, these models can provide valuable insights for the design of support systems and the evaluation of their performance in time (e.g. Sulem et al. 1987. Int J Rock Mech Min Sci Geomech Abstr, 24(3):145–154; Guayacán-Carrillo et al. 2016. Rock Mech. Rock Eng. 49(1):97-114; Liu, et al. WTC 2019). During the last decade, machine learning techniques have experienced vast growth in geotechnical engineering. These techniques present advantages concerning their computational performance and their applicability to high-dimensional non-linear problems. With the emerging use of these techniques, some questions arise as: (1) how these ones will contribute to the design of underground structures? and (2) what is its efficiency and accuracy using small datasets, as it is the case in rock engineering projects? The present work aims to propose a simplified Symbolic Regression (SR) approach in order to evaluate its applicability on the convergence evolution prediction. SR is a machine learning technique that aims to identify an underlying mathematical expression that best describes a relationship between input and output parameters (Stephens 2019. gplearn.readthedocs.io; Koza, J. (1992). MIT Press). In this work, special attention is given to anisotropic closure evolution, which depends on the anisotropy of the initial stress state and the intrinsic anisotropy of the rock mass formation. Moreover, the time needed on convergence monitoring is also studied, by taking into account different time intervals (corresponding to the duration of convergence monitoring). The results show that SR present a good predictive accuracy in comparison with the semi-empirical law proposed by Sulem et al. (1987) [Int J Rock Mech Min Sci Geomech Abstr 24(3): 145–154]. It is observed that this approach performs well with the small dataset used in this study and can be considered a useful alternative.

The rock mass properties are crucial for any structural design and the estimation of these properties is done using the Geological Strength Index (GSI). This study introduces a methodology to predict the GSI values using advanced image processing and Artificial Neural Network (ANN) techniques. The pictures and the face mapping data from an Iron ore mine are taken and using image processing techniques, including black-and-white conversion, joint highlighting, and noise reduction, the fractal dimensions of the rock faces are evaluated. An ANN model is developed which utilizes the fractal dimensions and the surface condition index as inputs and predicts the GSI values. This methodology aims to overcome the subjectivity of qualitative assessments, providing a more accurate representation of rock mass strength. The R2 value of the developed ANN model is 0.67 which indicates a positive correlation between predicted and qualitative GSI values from the standard chart.

The expansion of the underground infrastructure and the necessity to maintain the functionality of existing tunnels highlights the role of structural health monitoring to track the behavior of the structure. In the study performed, the focus is on segmental tunnel linings for deep and long tunnels. The aim is to reconstruct in real-time the stress field in the tunnel structure starting from a few monitoring data. The framework is based on the combination of finite element (FE) models of the lining in the hosting rock mass and machine learning algorithms, which are deployed for real-time estimation of the stress distribution in the lining. An approach based on synthetic data generated with FE simulations permits to reconstruct a thorough picture of the structural stress state based on a few monitored points. The method is applied to a full-scale test, in which three piled lining rings were tested under geostatic-like loads.

In NATM tunnel excavation sites, rock mass evaluation at tunnel face heavily influences the determination of support patterns of tunnel construction. Currently, this evaluation is performed by skilled engineers who score the rock mass according to several predetermined evaluation criteria. This method heavily relies on the subjective standards of the engineers. Thus, a uniformed and standardized evaluation method has not been established. In this study, the applicability of machine learning to rock mass evaluation is verified to achieve a qualitative evaluation method. A support vector machine (SVM) model was developed to predict the support pattern of one tunnel face to test applicability at tunnel construction sites.

We propose a new approach leveraging 3D Hough transform to detect discontinuity planes from traces in digital rock models, where clear planar surfaces are not apparent. Our method assumes that that the discontinuity plane behind a trace on the rock surface aligns roughly parallel to the trace's local direction, like a door rotating around a hinge. We begin by estimating local trace directions using multi-scale principal component analysis. These directions then inform the weighted voting of trace points within the Hough space. Peak-searching identifies initial planes, which undergo refinement to eliminate spurious or duplicate results. Applied to synthetic data and two natural outcrops, our method performed discontinuity extraction from traces with varying degree of success. This approach shows promise as an analytical tool in geology with potential for further optimization.