Machine Learning based Analysis of RVE in Concrete Structures

Resumo

In the numerical modeling of structural elements, particularly utilizing the finite element method (FEM), multiscale models represent the most robust approach for analyzing heterogeneous materials such as concrete. These models incorporate information from multiple scalestypically, macroscopic and mesoscopicto more accurately capture the material's response and its impact on structural behavior. In this context, each material point in the numerical model is associated with a Representative Volume Element (RVE), which is the smallest statistically representative portion of the material. During the analysis of a multiscale model via FEM, each integration point in the model requires the execution of an independent computational model that simulates the RVE and is processed in parallel to the main model (on the fly process). Deformations originating from the macroscale are applied to the mesoscale models, and the responses obtained are then reintegrated into the macroscale. This approach results in significant demand on processing resources and time, which can limit its practical application in large-scale projects or situations requiring rapid responses. This article propose the representation of the RVE through a Machine Learning model (ML) capable of simulating the mechanical behavior of the material. This approach eliminates the on the fly processing. Throughout the text, examples of RVEs represented by ML models are presented, and their specificities are discussed. Finally, the advantages and disadvantages of this methodology are examined.