Neural networks for defining spatial variation of rock properties in sparsely instrumented media

Reliable information of the three-dimensional distribution of rock mass properties improves the design of secure and cost-effective civil structures.In this paper, a recurrent neural network is presented as an alternative to predict the spatial variation of some index properties of rock in sparsely instrumented media.The neural technique, from statistical learning models, is used to approximate functions that can depend on a large number of inputs that are generally unknown.From a reasonably simple neuronal model of two inhomogeneous rock volumes, the limited measured information is extrapolated and the properties in the entire mass can be estimated.Comparisons between in situ explorations versus the 3D-neuronal definition confirm the potential of the proposed method for characterizing the properties of masses with inhomogeneous properties.Such a representation is useful for design of economic realistic numerical modelling of rock volumes, maximizing information while minimizing cost.