An artificial neural networks model for concrete in plasticity

Abstract

A data driven plasticity-oriented approach is developed to represent the nonlinear compressive response of plain concrete under uniaxial and biaxial loading. The proposed framework combines the constitutive concepts of plasticity with an artificial neural network (ANN), specifically a radial basis function (RBF) network, to approximate the stress-strain response over a range of compressive strength, elastic modulus, and biaxial confinement conditions. The ANN is trained and validated using experimental data reported in the open literature, with the available input-output pairs divided into training, validation, and test subsets. The network parameters are selected by minimizing the mean squared error on the validation data. The resulting model is assessed against independent experimental stress-strain responses for uniaxial and biaxial compression. The comparisons show that the RBF-based model reproduces the principal trends of the experimental responses, including nonlinear hardening, peak-strength behavior, post-peak softening, and the influence of lateral compression. The results indicate that an ANN can provide an efficient surrogate for describing complex compressive response while retaining a direct connection with experimentally observable materiel parameters. The study also identifies the main limitations of the original data-driven formulation, particularly its dependence on the experimental database and the need for broader multiaxial and loading-path validation. In this context, the proposed approach provides a useful foundation for the development of more general physics-informed and data-driven constitutive models for concrete.