Optimization of the concrete material parameters in numerical simulation of RC beams under shear failure by artificial neural networks

Resumo

Three-dimensional nonlinear constitutive models for reinforced concrete (RC) often require an
extensive number of material parameters. Some of these are evaluated through experimental tests, but most of
them are estimated from empirical or semi-empirical expressions. The inherent limitations of representing real
structures with mathematical models may introduce different constraints on the parameter calibrations that
originally adjusted those expressions. That is one of the main reasons that lead to divergences between nonlinear
finite element analysis (NLFEA) and reliable experimental data. In this study, we employ a method that adopts an
artificial neural network (ANN) and Levenberg-Marquatd backpropagation method for calibrating material
parameters in a numerical model of an RC member. The simulated experiment is an RC beam under a three-point
bending scheme with shear failure. Finite element computations are carried out in ATENA software, and the goal
of the calibration is to find the best adjustment of the experimental load-deflection curve at the center of the beam.
The algorithm shows an excellent capability to fit the numerical curves, and it successfully automates a task that
customarily would take several analyses and trial-and-error process to achieve the best fitting.