Dataset for “Interpretable ANN-Based Constitutive Modeling for Quasi-Brittle Materials from Synthetic FEM Data”
Description
This dataset contains synthetic stress–strain histories generated from nonlinear finite element simulations and used for the development and assessment of an Artificial Neural Network-based Constitutive Model (ANN-CM) for quasi-brittle materials, with concrete adopted as the representative material. The repository includes numerical data from several structural and material configurations, including plate simulations under different uniaxial, biaxial, shear, and tension-compression stress states, as well as bending, shear, diametrical compression, pre-stressed bending, asymmetric traction, and L-shaped panel tests. Selected plate simulations were also performed with different rotation angles to expand the range of stress-state orientations represented in the database. The data are organized into training and independent assessment subsets. Each record corresponds to an integration-point state and contains the current stress–strain state, two previous stress–strain states, the imposed strain increment, and the corresponding stress increment used as the ANN-CM training target. Under plane-stress conditions, the learning problem comprises 21 input variables and 3 output stress-increment variables. The repository also includes a README, a data dictionary, and a file manifest describing the variable definitions, units, simulation naming conventions, and the role of each file. These data support the study “Interpretable ANN-Based Constitutive Modeling for Quasi-Brittle Materials from Synthetic FEM Data” and the associated M.Sc. dissertation “The challenges of concrete constitutive modeling via artificial neural networks.”
Files
Steps to reproduce
The dataset was generated from nonlinear finite element simulations performed in INSANE (Interactive Structural Analysis Environment), developed at the Department of Structural Engineering of the Federal University of Minas Gerais, Brazil. Concrete behavior was represented using a fixed-crack smeared constitutive model with the stress–strain relationships proposed by Carreira and Chu for compression and tension. Material properties, numerical parameters, geometries, boundary conditions, loading procedures, and convergence criteria for the simulations are described in the associated manuscript and M.Sc. dissertation. Several numerical configurations were analyzed to generate different stress–strain states and loading histories, including plate simulations under uniaxial tension and compression, biaxial tension and compression, combined tension–compression, and pure shear, as well as three-point bending, four-point bending, shear, diametrical compression (Brazilian splitting), pre-stressed bending, asymmetric traction, and L-shaped panel tests. For selected plate simulations, the element/reference system was rotated through different angles to generate additional orientations of the stress state. Stress and strain histories were extracted at finite element integration points for each equilibrium step. For every material state, the exported data contain the current stress and strain components, two previous stress–strain states, the imposed strain increment, and the corresponding stress increment. These quantities were subsequently organized to define the 21 ANN-CM input variables and the three stress-increment target variables described in the manuscript. Training and assessment data were separated at the simulation-file level. Structural configurations designated for independent assessment were excluded from the training dataset. More than 90 plate simulation cases were generated; after excluding the held-out plate cases, a subset of the remaining plate-derived data was selected to avoid overrepresentation of plate simulations relative to the other structural configurations. During preparation of the ANN-CM training database, the variables were normalized to the range [-1, 1]. Additional intermediate states were generated by linear interpolation between consecutive equilibrium states to increase the variability of the strain increments represented during training. The original finite-element-derived histories provided in this repository, together with the manuscript, the accompanying README, DATA_DICTIONARY.csv, and FILE_MANIFEST.csv, document the dataset structure and the procedure used to reconstruct the learning problem. Full methodological details are available in: Figueiredo, A. F. (2023). The challenges of concrete constitutive modeling via artificial neural networks. M.Sc. Dissertation, Federal University of Minas Gerais, Belo Horizonte, Brazil.
Institutions
- Universidade Federal de Minas GeraisMinas Gerais, Belo Horizonte