Rana Magar, Bibek (2025) Surrogate Models for Improved Performance in Structural Safety Performance Assessment. PhD thesis, Victoria University.

Abstract

This thesis is concerned with a model surrogation framework whereby the performance in certain structural safety assessment processes can be improved. In particular, the study aims to develop suitable surrogate models to reduce the time spent on numerical or laboratory simulation components in the process of Monte Carlo simulation-based reliability and safety assessment. A comprehensive review of structural safety assessment methodologies was conducted to establish the need for surrogate modelling. The review highlighted those conventional approaches, including Monte Carlo simulation, Bayesian methods, and time-marching algorithms, often require thousands of simulations runs and substantial computational resources, limiting their practical application in engineering design and assessment. It is important to replace these time-consuming components with suitably constructed surrogate models. To fulfill the scope of the study, a numerical study was conducted to examine a two-storey nonlinear frame structure subjected to strong wind loads. The fragility analysis of the structure was assessed using the inter-storey displacement as the primary performance. Due to the highly variable and dynamic nature of wind loads, structural response analysis can be computationally demanding Therefore, force time histories and storey stiffness parameters were used as input variables to develop a neural network-based surrogate model capable of predicting inter-storey displacements. It was demonstrated that well-constructed surrogate models can accurately reproduce the simulation results of classical time-marching algorithms while substantially improving computational efficiency and significantly reducing computational cost. The framework was subsequently extended to the fragility assessment of a three-storey building subjected to earthquake actions. Nonlinear structural analyses were performed using industry-standard design software to account for material failure under seismic actions. Variations in material properties were incorporated to generate datasets containing inter-storey and horizontal displacement responses. These datasets were used to train neural network based surrogate models that effectively reproduced the results of the nonlinear simulations, providing a computationally efficient alternative to conventional analysis procedures. The research further examined the development of surrogate models for analytical and laboratory-based simulations involving storage silos and backfilled mine stopes. Understanding stress development in these systems is essential for ensuring structural safety and operational reliability. An initial pilot study based on published experimental data facilitated the development of a preliminary surrogate model for silo structures. This was followed by laboratory experiments on a silo prototype and backfilled mine stope systems. Vertical stresses at the silo base and shear stresses along the silo wall were measured under varying filling materials, fill heights, and wall roughness conditions. Experimental data were subsequently used to train neural network surrogate models capable of accurately predicting stress distributions in both silos and backfilled mine stopes. The accuracy of the developed neural-network based surrogate models was evaluated by comparing their predictions against results derived from the traditional method (numerical simulation/ experimental method). In all cases, the surrogate models produced results within acceptable error limits, demonstrating their reliability as replacements for more computationally demanding assessment methods. Finally, the applicability of the developed surrogate models within complete structural safety assessment frameworks was investigated. The earthquake and silo surrogate models were integrated into reliability analysis to estimate the probability of failure of representative structures under prescribed loading conditions. Failure criteria were defined using threshold values of inter-storey displacement and stress, respectively. Monte Carlo simulations involving thousands of simulation runs were performed using the surrogate models, enabling rapid estimation of failure probabilities with substantially reduced computational time. The findings confirm that neural network-based surrogate modelling provides an effective and efficient approach in reducing the time consumed by classical analysis techniques while maintaining acceptable levels of accuracy The study also identifies opportunities for future improvements through the incorporation of more advanced input variables and modelling strategies, enabling the application of surrogate modelling to increasingly complex real-world engineering problems.

Search Google Scholar

Repository staff login