Evaluation of previous studies of multi-layer perceptron type artificial
neural networks in the modelling of the deformability behaviour of rock
have suggested that a deviation from the standard method of laboratory
testing of rock under triaxial loading at constant confining stresses and
constant axial strain rates may be inappropriate when the data produced
are used to train an artificial neural network. In this paper, preliminary
results of adopting such a strategy are reported. Of specific concern is the
establishment of a suitable schedule for the stress path during testing.
Furthermore, the training element of back propagation optimisation was
directed towards prediction of new stresses given old stresses and old
strain increments. Adoption of this method of supervisation in artificial
neural network training facilitates the integration of the trained neural
network within existing numerical methods used in rock mechanics. The
target code was a finite difference Lagrangian formulation for the analysis
of continua. Sandstone rock taken from a quarry in the UK was used in
the study which showed that the method has great potential.