TY - CHAP A1 - Unger, Jörg F. A1 - Könke, Carsten ED - Gürlebeck, Klaus ED - Könke, Carsten T1 - PARAMETER IDENTIFICATION OF MESOSCALE MODELS FROM MACROSCOPIC TESTS USING BAYESIAN NEURAL NETWORKS N2 - In this paper, a parameter identification procedure using Bayesian neural networks is proposed. Based on a training set of numerical simulations, where the material parameters are simulated in a predefined range using Latin Hypercube sampling, a Bayesian neural network, which has been extended to describe the noise of multiple outputs using a full covariance matrix, is trained to approximate the inverse relation from the experiment (displacements, forces etc.) to the material parameters. The method offers not only the possibility to determine the parameters itself, but also the accuracy of the estimate and the correlation between these parameters. As a result, a set of experiments can be designed to calibrate a numerical model. KW - Angewandte Informatik KW - Angewandte Mathematik KW - Architektur KW - Computerunterstütztes Verfahren KW - Computer Science Models in Engineering; Multiscale and Multiphysical Models; Scientific Computing Y1 - 2010 U6 - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:gbv:wim2-20170314-28984 UR - http://euklid.bauing.uni-weimar.de/ikm2009/paper.html SN - 1611-4086 ER - TY - CHAP A1 - Schrader, Kai A1 - Könke, Carsten ED - Gürlebeck, Klaus ED - Könke, Carsten T1 - SPARSE APPROXIMATE COMPUTATION OF SADDLE POINT PROBLEMS ARISING FROM FETI-DP DISCRETIZATION N2 - The numerical simulation of microstructure models in 3D requires, due to enormous d.o.f., significant resources of memory as well as parallel computational power. Compared to homogeneous materials, the material hetrogeneity on microscale induced by different material phases demand for adequate computational methods for discretization and solution process of the resulting highly nonlinear problem. To enable an efficient/scalable solution process of the linearized equation systems the heterogeneous FE problem will be described by a FETI-DP (Finite Element Tearing and Interconnecting - Dual Primal) discretization. The fundamental FETI-DP equation can be solved by a number of different approaches. In our approach the FETI-DP problem will be reformulated as Saddle Point system, by eliminating the primal and Lagrangian variables. For the reduced Saddle Point system, only defined by interior and dual variables, special Uzawa algorithms can be adapted for iteratively solving the FETI-DP saddle-point equation system (FETI-DP SPE). A conjugate gradient version of the Uzawa algorithm will be shown as well as some numerical tests regarding to FETI-DP discretization of small examples using the presented solution technique. Furthermore the inversion of the interior-dual Schur complement operator can be approximated using different techniques building an adequate preconditioning matrix and therewith leading to substantial gains in computing time efficiency. KW - Angewandte Informatik KW - Angewandte Mathematik KW - Architektur KW - Computerunterstütztes Verfahren KW - Computer Science Models in Engineering; Multiscale and Multiphysical Models; Scientific Computing Y1 - 2010 U6 - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:gbv:wim2-20170314-28874 UR - http://euklid.bauing.uni-weimar.de/ikm2009/paper.html SN - 1611-4086 ER - TY - CHAP A1 - Ahmad, Sofyan A1 - Zabel, Volkmar A1 - Könke, Carsten T1 - WAVELET-BASED INDICATORS FOR RESPONSE SURFACE MODELS IN DAMAGE IDENTIFICATION OF STRUCTURES T2 - Digital Proceedings, International Conference on the Applications of Computer Science and Mathematics in Architecture and Civil Engineering : July 04 - 06 2012, Bauhaus-University Weimar N2 - In this paper, wavelet energy damage indicator is used in response surface methodology to identify the damage in simulated filler beam railway bridge. The approximate model is addressed to include the operational and surrounding condition in the assessment. The procedure is split into two stages, the training and detecting phase. During training phase, a so-called response surface is built from training data using polynomial regression and radial basis function approximation approaches. The response surface is used to detect the damage in structure during detection phase. The results show that the response surface model is able to detect moderate damage in one of bridge supports while the temperatures and train velocities are varied. KW - Angewandte Mathematik KW - Computerunterstütztes Verfahren KW - Angewandte Informatik Y1 - 2012 U6 - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:gbv:wim2-20170306-27588 SN - 1611-4086 ER -