Carsten Könke

An inverse parameter identification procedure assessing the quality of the estimates using Bayesian neural networks

Abstract: The paper proposes a general procedure based on Bayesian neural networks for parameter identification of numerical models. In this context, the Bayesian neural networks are extended to multiple outputs with a full covariance matrix to describe the correlation between the noise of output parameters. This extension is especially useful for inverse problems such as a parameter identification procedure, since it allows for the quantification of correlations between output parameters. Based on numerically obtained forward calculations, the Bayesian neural network is trained to solve the inverse parameter identification problem. The main advantage of the method is the ability to verify the accuracy of the identified parameters and their correlation. The methodology further allows to detect, whether a certain set of experiments is sufficient to determine an individual model parameter. As a result, a general scheme for the design of experiments to identify model parameters is developed and illustrated for two examples.

Blast Demolition Study of Guyed Masts

Demolition of guyed masts is usually carried out by cutting down some of the supporting guy cables using an explosive in such a way that the mast can fall into the desired direction. Without the cable supports, guyed tubular masts are very slender structures which are susceptible to local buckling based on the internal force distribution. If this local buckling occurs at the early stage of the demolition processes, it can cause uncertainty in the failure mechanism. The risk of undesirable demolition outcome due to this uncertainty can be mitigated by using controlled detonation setups. In this paper, a sensitivity analysis is presented using a case study to determine the influence of the explosive detonation time on the collapse development and pattern of the guyed mast. Then, the results of the sensitivity analysis are systematically categorized using cluster analysis to show possible types of collapse regimes which can be used to setup a controlled demolition scheme.

Coupling between Shell and Generalized Beam Theory (GBT) elements

This work presents a procedure to couple shell and Generalized Beam Theory, GBT, elements. The main focus of this procedure is the possibility to model mixed beam frame structures, which the traditional shell elements are applied at the joints and GBT elements are used to model the beams/columns. Such modeling technique can use the benefits of both elements. At the joints, shell elements can easily simulate different types of geometry conditions and details, such as stiffeners and holes; meanwhile, for the beams and columns, GBT can provide high performance, accuracy and an easy modeling approach with clear results.

On the Disproportionate Collapse Risk and Robustness in the Eurocode

After the partial collapse of Roman Point residential tower in London in 1968, the tensile catenary, or the known tie-forces’ method, is the norm for robustness check of building in risk classes A and B2 of the Eurocodes. This method is discussed here in light with recent developments in testing and simulation. It is shown that this principle can be further improved, for example it can be proven that the tie-forces are inefficient for the corner assemblies, and without sufficient lateral stiffness, it increases the risk of pull-down of other parts of the structure.

Energy Methods for Assessing Dynamic SSI Response in Buildings

An energy approach is employed here for assessing model quality for dynamic soil-structure interaction (SSI) analysis. Concurrently, energy measures are introduced and calibrated as general indicators of structural response accuracy. More specifically, SSI models built at various abstraction levels are investigated according to various coupling scenarios between soil and structure. The hypothesis of increasing model uncertainty with decreasing complexity is first investigated. A mathematical framework is provided, followed by a case study involving alternative models for incorporating SSI effects. During the evaluation process, energy measures are used in conjunction with the adjustment factor approach to quantify SSI model uncertainty. Two types of uncertainty are considered, namely in the numerical model and in the model input parameters. Investigations on model framework uncertainty show that the 3D finite element (FE) model yields the best quality results, whereas the Wolf lumped parameter model produces the lowest model uncertainty in the simple model category. Also, the fixed-base model produces the highest estimated uncertainty and consequently is the worst quality model. The present results confirm the hypothesis that increasing model uncertainty comes with decreasing complexity, but only when the assessment is based on an energy measure as the response indicator.

Distributed FE Analysis of Multiphase Composites for Linear and Nonlinear Material Behaviour

Modern and efficient adaptive multiscale models can be applied for the prognosis of complex material behavior including deterioration and damage effects of heterogeneous materials. Therewith the physical effects of damage initiation and crack propagation can be captured on the appropriate spatial scales and parameter identification for the different material constituents can be performed much easier than by applying phenomenological material models for composite materials on macroscale. As a consequence of application of these models, the number of degrees of freedom and therewith the necessary computing effort increases substantially. This statement holds for the necessary main memory as well as for the computing power to solve the underlying linear and nonlinear equation systems. The project dcmamc have been investigating partition/substructuring methods and efficient parallel algorithms to solve very large linear and nonlinear equation systems. The algorithms were implemented, tested and adapted to the HPC computing framework at the HLRS Stuttgart, using several hundred CPU nodes. A memory-efficient iterative and parallelized equation solver combined with a special preconditioning technique for solving the underlying equation system was modified and adapted in order to be applied in a mixed CPU and GPU based hardware environment. Additionally a saw-tooth algorithm has been adapted to take into account nonlinear material behavior in a sequential linear manner. Therewith the material nonlinear problem is treated as a sequential solution of purely linear problems, avoiding all drawbacks with convergence problems in classical incremental-iterative solution techniques. In return a substantially increased number of linear solution steps has to be accepted.

Distributed FE Analysis of Multiphase Composites Regarding 3D Elasticity Problems

Today the numerical simulation of damage effects in heterogeneous materials is done by adaption of multiscale approaches. A consistent modeling in three dimensions with a high discretization resolution on each scale based on a hierarchical or concurrent multiscale model still has issues. The algorithms have to be optimized in regards to the computational efficiency and the distribution among available hardware resources often based on parallel hardware architectures. In the last 5 years high performance computing (HPC) as well as GPU computation techniques were established for investigation of scientific aims. Consequently, in this work substructuring methods for partitioning of FE meshed specimens were implemented, tested and adapted to the HPC computing framework using several hundred CPU nodes. An memory-efficient iterative and parallelized equation solver combined with a special preconditioning technique for solving the underlying equation system was modified and adapted to the consideration of combined CPU and GPU based computations.

Analysis of Crack Initiation and Propagation in Polycrystalline Meso and Microstructures of Metal Materials

Durability and life cycle analysis of engineering structures is often based on numerical simulations of macroscopic damage behaviour using phenomenological damage and fracture models. Therewith the true mechanisms of crack initiation and various crack propagation can not be covered. In order to integrate the physical material effects which are leading to crack initiation as well as crack propagation, simulations on the meso- or microstructure have to be performed.