Abstracts

Utilização de simulações numéricas, inteligência artificial e algoritmos de amostragem inteligente para construir modelos substitutos e calcular a probabilidade de falha de túneis urbanos: Combining numerical simulations, artificial intelligence and intelligent sampling algorithms to build surrogate models and calculate the probability of failure of urban tunnels

by Vinícius Resende Domingues

When it is necessary to evaluate, with a probabilistic approach, the interaction of urban tunnels with neighboring structures, computational power is an important challenge for numerical models. Thus, intelligent sampling algorithms can be allies in obtaining a better knowledge of the result domain, even if in possession of a smaller number of samples. In any case, when sampling is limited, the evaluation of the building risks is also restricted. It is in this context that artificial intelligence can fill an important gap in risk analysis by interpolating results and generating larger samples in a short time. In this thesis a hypothetical case was used to validate the methodological proposal. It concerns the sequential excavation, as the NATM, of a tunnel, three-diameter deep, interacting with a building containing seven floors. First, the threedimensional numerical model (FEM) was solved deterministically, and then its domain and mesh were refined. After that, another 170 solutions were numerically obtained from a FEM software, strategically sampling the random variables involved. Sequentially, based on the 31 artificial intelligence techniques, it was evaluated which variables were of greatest importance to predict the magnitude of vertical displacement in the foundation elements of a surrounding building. Then, once the most important variables were selected, the 31 artificial intelligence techniques were again trained and tested to define the one with the least R-squared. Finally, by using this best-fit algorithm, it was possible to perform the calculation of the probability of failure using massive samples, with sizes on the order of 107 . These samples were used to check the convergence of simple Monte Carlo sampling and its variations, as well as the semianalytical FOSM, FORM, and SORM methods. The main contribution of this thesis is methodological; therefore, this new procedure can be aggregated to state-of-the-art risk assessment methodologies in tunnel-related problems.