Surrogate modelling for finite element simulations Towards a stochastic surrogate approach

To reduce required used knowledge and simultaneously improve generalization in finite element surrogate modelling, a stochastic weight assembly is introduced to the well-known traditional long short-term memory (LSTM) neural network. Due to the propagation of uncertainties and simultaneous sparsity favouring weight optimization, it is shown using a structural application that an auto-adaptive recurrent LSTM neural network is created which automatically adapts its architectural complexity to the problem provided, hereby improving LSTMs generalization capabilities.