Machine learning based seismic stability assessment of dams with heterogeneous concrete

Abstract

Nowadays, the probabilistic risk assessment (PRA) is one of the standard tools for safety evaluation of the existing dams. Although many factors are involved in risk-informed condition assessment of infra-structures, a detailed numerical simulation plays the key role. A comprehensive numerical model usually includes the interaction of the dam with foundation and reservoir, nonlinearity of the concrete, and appropriate seismic wave propagation. Often, the concrete is assumed to be a homogenous material which facilitates the transient simulations. However, in reality, the mass concrete has a heterogeneous nature especially in the case of old dams suffering from aging and deterioration. The objective of this joint contribution is to evaluate the seismic response of a typical gravity dam with heterogeneous concrete material. Since the simulations are cast in the context of probabilistic methods, a relatively large number of analyses are required to fully understand the impact of concrete random field on the structural responses (i.e. displacements, and stresses). Therefore, a Random Forest (RF) meta-model is adopted in order to reduce the computational burden of the probabilistic simulations. RF is an effective technique in the regression analyses of the systems with large number of inputs, and can provide a measure of input variables' importance. This technique is useful to identify the critical areas within the dam body which deterioration may affect the seismic stability of the dam and jeopardize its integrity.