PREDICTING THE NATURAL FREQUENCIES OF TRUSS DOMES UNDER UNCERTAINTY USING DEEP FEEDFORWARD NEURAL NETWORKS

This study employs Monte Carlo simulation together with a deep feedforward neural network to predict the natural frequencies of truss domes under uncertainty. Material and/or geometric properties of these structures are modeled as random variables, and their influence on the natural frequencies is examined. Monte Carlo simulation is applied to perform stochastic eigenvalue analyses of the finite element models. To reduce computational cost, a deep neural network is trained to predict natural frequencies in place of repeated eigenvalue solves, accelerating the overall simulation. Bayesian optimization is used to tune the network hyperparameters. Numerical examples show that the proposed approach substantially improves computational efficiency and predictive accuracy compared with direct Monte Carlo simulation for domes with random inputs.

Keywords:

Machine learning , feedforward neural network , Dome , Random eigenvalue problem , Bayesian optimization , Monte Carlo simulation