Immo LUKAS
Application of high-performance computing for the seismic collapse assessment of a planar steel frame
structure
Natural hazards have always been among the most dangerous events for humanity. Despite all technological progress, recent seismic events have demonstrated their destructive power. Particularly in an urban environment, handling the effects of earthquakes on buildings and infrastructure remains a core responsibility for the engineering community. In earthquake engineering, seismic collapse assessment is among the most important concerns in the evaluation of a seismically excited structure. Non-linear dynamic analyses provide the opportunity to realistically predicting building behavior under seismic excitation. These analyses are computationally demanding and, therefore, rarely used in practice. For this reason, as part of a large-scale study of the Unit of Applied Mechanics of the University of Innsbruck, a database is generated that yields the calculations' results of multi-storey frame-like structures and thus provides the basis for a deeper understanding of their collapse behavior under earthquake excitation. The long-term goal is to sidestep the computationally expensive calculations for seismic collapse assessment with machine learning-based predictions. Based on the NGA-West2 ground motion database, 17 000 ground motions are used for Incremental Dynamic Analysis. Within the scope of this Master thesis, an 8-storey moment-resisting steel frame structure is considered. The introduction includes a composition of the necessary fundamentals of earthquake engineering and provides insights regarding the modeling and implementation of the steel frame structure. The evaluation of the results reveals a significant shift of the deformation pattern in the collapse limit state compared to the evaluation point just previously to collapse. Common intensity measures are assessed according to the criteria of efficiency, sufficiency, and scalability. When examining the influence of individual storeys on the overall collapse behavior, it is observed that particularly the collapse of upper storeys strongly correlates with a high scale factor, which in turn distorts the ground motion record and thus presumably provides limited information. Finally, the machine learning algorithm K-means clustering aims to structure the database of ground motion records according to similar characteristics and group them into data sets. Based on the intentionally selected sets, the prediction accuracy of existing generalized linear machine learning models is supposed to be increased.