A polyaxial strength model for intact sandstone based on Artificial Neural Network

A comprehensive database of the polyaxial compressive strength of eight sandstones has been compiled from the literature. An experimental study has also been conducted on local sandstone to add up to the database. A correlation-based analysis has been performed to find out the influence of each independent parameter namely uniaxial compressive strength (UCS, sigma(ci)), minor principal stress (sigma(3)), and intermediate principal stress (sigma(2)) on the strength of sandstone, i.e., the major principal stress at failure (sigma(1)). Additionally, a feed-forward back propagating neural network (FFBPNN) has been proposed as a new polyaxial strength model to predict the strength of intact sandstone under polyaxial states of stresses. The database on polyaxial strength of 192 experiments on nine different sandstones has been randomly divided into a training set and a testing set. Three input parameters corresponding to the independent parameters (sigma(ci) sigma(3), sigma(2)) and one output parameter corresponding to the dependent parameter (sigma(1)) are considered. The accuracy of the proposed ANN based polyaxial strength model has been compared with five other conventional polyaxial criteria: modified Wiebols and Cook criterion (MWC), Mogi-Coulomb criterion (MC), modified Lade criterion (ML), 3D version of Hoek Brown criterion (3DHB) and modified Mohr-Coulomb criterion (MMC). It is found that the ANN based failure model gives the best result amongst all the considered polyaxial strength criteria, for the testing dataset.