Numerical-Upscaling-Based Dataset for Predicting the Effective Stiffness Matrix of Fluid-Filled Microcrack-Bearing Rocks

This record provides the dataset used to train and evaluate a numerical-upscaling-based multi-task learning multilayer perceptron (MTL-MLP) model for predicting the full effective stiffness matrix of heterogeneous rocks containing fluid-filled microcracks. The dataset was generated from physics-based numerical upscaling simulations based on coupled solid-fluid formulations and oscillatory relaxation tests. Each sample links four dimensionless crack descriptors to six effective stiffness components. The input variables are crack density (CD), aspect ratio (AR), crack-spacing descriptor (Hv), and crack-length-to-sample-size ratio (CL/W). The target variables are the effective stiffness coefficients C11, C12, C22, C66, C16, and C26. This dataset is intended to support reproducible research on data-driven surrogate modeling, intelligent homogenization, and rock-physics prediction in fractured fluid-bearing rocks. The record includes the raw or processed data files used for model development, together with the metadata needed to interpret the variables and reproduce the learning workflow. Users of this dataset are requested to cite both this Zenodo record and the associated journal article when using the data in publications or derivative works.