Coupling machine learning and stochastic finite element to evaluate heterogeneous concrete infrastructure

Abstract

Probabilistic analyses facilitate the incorporation of inherent uncertainties in dam safety evaluation. Their application is limited due to the high computational burden, especially for the complex transient analyses and the cases in which the random field theory is applied to account for spatial uncertainty in material properties. In this work, uncertainty in the heterogeneous concrete is modeled by a series of probabilistic seismic finite element simulations. The outcome of stochastic models is fed into a random forests model for further post-processing and sensitivity analysis. The challenging aspect of this research lies in the fact that the initial number of finite element realizations is relatively small (due to computational time) regarding the random variables. Despite such an unfavorable setting, the meta-model showed reasonable accuracy for predicting the maximum displacement at the crest and the maximum stress at the base. A comprehensive discussion is provided on the variable importance and the partial effect of the inputs: the areas in the dam body with the highest impact in the response are identified, and the spatial distribution of input material properties is appropriately captured. These results show the possibilities of random forests in high-dimensional problems for predicting and interpreting dam behavior.