From Gaussian to lognormal: improving material property modeling for precise structural predictions

Accurate prediction of material properties is essential in structural engineering design to ensure reliability and safety. Traditional approaches often rely on Gaussian distributions to model material variability. However, our research reveals limitations with Gaussian assumptions, particularly when covariance parameters exceed certain thresholds, leading to physically unrealistic negative values for material properties. To overcome these limitations, we investigate an alternative approach using lognormal distributions for material property modeling. Through Monte Carlo simulations employing Cholesky decomposition, we compare the performance of lognormal distributions with Gaussian counterparts. Our findings demonstrate that lognormal distributions offer a viable alternative, providing more accurate representations of material variability while maintaining computational efficiency. Furthermore, we utilize finite element method (FEM) data from Monte Carlo simulations to predict beam deflection using deep neural networks (DNNs). Leveraging inverse modeling techniques, we showcase the ability to predict elastic modulus from beam deflection data under both normal and lognormal distribution assumptions. By integrating lognormal modeling and inverse modeling techniques into structural analysis, we enhance the realism and accuracy of predictions, thereby improving reliability in engineering design. This paper discusses the implications of our findings and emphasizes the importance of considering alternative probability distributions for material property modeling in structural engineering applications.