Knowledge‐Based Modeling of Material Behavior with Neural Networks

TLDR

Material modeling has traditionally relied on mathematical models derived from human observation and reasoning with experimental data. The paper proposes using neural networks to model material behavior, specifically applying back‑propagation networks to concrete under plane stress and cyclic loading. Neural networks learn stress–strain relationships directly from experimental data within a unified network, as demonstrated with back‑propagation models applied to concrete under biaxial and cyclic loading. Preliminary results indicate that neural‑network modeling offers promising, scalable insights for complex materials such as composites.

Abstract

To date, material modeling has involved the development of mathematical models of material behavior derived from human observation of, and reasoning with, experimental data. An alternative, discussed in this paper, is to use a computation and knowledge representation paradigm, called neural networks, developed by researchers in connectionism (a subfield of artificial intelligence) to model material behavior. The main benefits in using a neural‐network approach are that all behavior can be represented within a unified environment of a neural network and that the network is built directly from experimental data using the self‐organizing capabilities of the neural network, i.e., the network is presented with the experimental data and “learns” the relationships between stresses and strains. Such a modeling strategy has important implications for modeling the behavior of modern, complex materials, such as composites. In this paper, the behaviors of concrete in the state of plane stress under monotonic biaxial loading and compressive uniaxial cycle loading are modeled with a back‐propagation neural network. The preliminary results of using neural networks to model materials look very promising.