Journal of the Korean Society of Mineral and Energy Resources Engineers ISSN:2288-0291(Print) 2288-2790(Online) 한국자원공학회지

A strategy to introduce artificial neural networks(NN) in finite element(FE) codes used for load- deformation analyses for rock structures has been proposed. At the heart of these types of codes is a pre-trained neural network based constitutive model(NNCM) plugged in a conventional FE program which can be used as usual to analyse structures and make predictions of deformations. For the NNCM based codes, the notion of material parameters is irrelevant and constitutive matrix is not necessarily required. However, although accuracy of non-linear FE computation can be achieved using an arbitrary constitutive matrix, it takes relatively long time to converge unless the constitutive matrix is known. There has been no way to obtain it from a trained NN. In this context, this paper provides a general formulation of the partial derivative of a NN. This enables us to compose the constitutive matrix (DNN) from a trained NNCM which can be plugged into a FE engine. An illustrative example is presented which shows that convergence of the NNCM based FE analysis is faster when the DNN is employed.