Maijia Su, Ph.D. Candidate 37th cycle, University of Trento, DICAM

The uncertainty in earthquake engineering is ubiquitous due to imperfect measurements, scarce data and the irreducibly intrinsic variability. In this scenario, a central task refers to quantifying the seismic risk so as to ensure life safety and control financial losses. This task requires two critical aspects, i.e., (I) identify and model the source of uncertainty, and (II) propagate the uncertain source to the quantities of interest (Figure 1 illustrates an example). The former generally tackle the randomness based on probability theory and calibrates the probabilistic models from collected data. The latter was the primary focus in the first Ph.D. academic year in which we developed the advanced computational tool for forward uncertainty quantification (UQ) analysis. Specifically, we investigated the active learning (AL) based surrogate modelling which allows seeking an inexpensive approximation model to substitute the computationally intensive simulator (e.g., the finite element solver) within an acceptable and minimum computational burden. We started the investigation by designing a generalized framework for constructing global surrogates to compute the full probability distribution. This framework includes three independent modules, i.e., types of surrogate models, design of experiment, and stopping criteria. Then we conducted a comparative study by choosing several representative methods for each module. 
In the following academic years, we will focus on modelling the uncertainty sources in earthquake engineering. We achieve this goal by using the copula theory which can capture complex dependencies among the random variables. Our final goal aims at building a holistic probabilistic network to connect all the random variables involved in seismic sources, soil conditions, seismic waves, building structures and structure seismic responses.

Figure 1: A illustration of forward uncertainty quantification that the probabilistic information propagates from the input variables to the outputs through the physics-based simulator.