From geometry to geomechanics: 2D bimrock AI-supported characterisation for underground engineering

This dissertation aims to better understand how the geometry seen on the field can be related to the geo-mechanical properties of rock masses composing the formations. Bimrock can present a huge obstacle when found due to their chaotic geo-mechanical behaviour. Composed by a fine ductile matrix interbedded with more competent fragile blocks they are known for not following neither classic rock mechanics laws nor soil laws. Different approaches were studied through the years, but numerical modelling proved to be the most effective even if a large amount of data was required. Due to the problematic application of this procedure in rock excavation processes it was decided to investigate if geo-mechanical properties present a correlation with geometric properties with simulation analysis. A neural machine was created and trained with 2D samples of Bimrocks (block-in-matrix rocks) subjected to a numerical simulation of compressed tests to gather the necessary parameters such as cohesion (MPa), friction angle (°), and the value of σ₁ (MPa) at 10MPa. The main differences between the samples were the orientation and blocks size. Another important diversity was the percentage of blocks inclusions (or volumetric block portion, VBP) inside the cylinders. Overall, three hundred models were created divvied up in ten classes. The first half of the samples (1-150) were created changing the percentage of block inclusions setting it at 10%, 25%, 45%, 60% and 80%. The second array of data (151-300) had a percentage of block inclusions set constant to 40%, but the parameter that was changed was their geometric properties. The maximum value for width and height of blocks was kept at 0.12u, the minimum had the following values 0.002u, 0.004u, 0.006u, 0.008u and 0.01u. This vast array of data was considered sufficient to train the neural machine. The output data confirmed that most parameters taken in set of two had not been related. It was discovered that some input parameters had a higher weight than others when confronting the output data found by the neural machine and the ones obtained with the compressed simulations test in FLAC. Inclusions percentage, area fraction, and distance from the centre of mass along the y axis were discovered to be the most important ones.