Many studies have worked on the estimation of fluid saturation as an important petrophysical property in hydrocarbon reservoirs. Based on Archie's law, proper determination of cementation factor (m) can lead to accurate values of water saturation. Given that the m is mainly affected by electrical properties of rock, electrical quality index (EQI) can be used to estimate m through a novel rock typing technique. Despite the efficient applicability of EQI for the classification of rocks, with similar electrical behaviors, into distinct electrical rock types (ERTs), manual implementation of this method is time-consuming and gets excessively more difficult for larger datasets. In this work, a fast automated version of EQI methodology was presented. As a fuzzy clustering algorithm, Gaussian mixture model (GMM) was implemented on a large quantity of carbonate and sandstone samples to cluster them into distinct ERTs based on EQI values. To this end, 100 data points were randomly selected for testing purposes, and the remaining data points were used as training subsets for carbonate and sandstone samples. An innovative hybrid EQI-GMM approach was developed to determine the optimum number of clusters. Furthermore, results of two commonly-used criteria, namely Schwarz's Bayesian Criterion (BIC) and Akaike Information Criterion (AIC), showed that they fail to specify ERTs properly. The predicted values for m by the hybrid EQI-GMM approach were more accurate (RMSE is 0.0167 and 0.0056 for carbonate and sandstone samples, respectively) than outputs of the traditional Archie’s law (RMSE is 1.6697 and 0.1850 for carbonate and sandstone samples, respectively).
4 October 2023 in Bulletin of Engineering Geology and the Environment
This study presents a graph neural network (GNN)-based spatio-temporal surrogate model for long-term prediction of heat transport in heterogeneous geothermal reservoirs. The proposed approach serves as a fast proxy for the conventional reservoir simulators. It combines local message passing on the simulation grid with physics-aware spatial tokenization, in which algebraic multigrid (AMG) coarsening concentrates representational capacity along the flow corridor connecting the wells. Temporal evolution is learned on the coarse graph through streamline-causal attention and a node-wise recurrent memory, rolling out autoregressively over a 100-year operational horizon. A dataset of 600 two-dimensional synthetic geothermal reservoir simulations is generated using the MATLAB Reservoir Simulation Toolbox (MRST), incorporating heterogeneous porosity and permeability fields, incompressible fluid flow, and coupled conductive--advective heat transport in thermal doublet configurations with varying well placements and injection rates. Comparison with high-fidelity numerical simulation on unseen cases shows that the model reproduces both the evolving temperature field and the produced-fluid temperature decline with sub-Kelvin accuracy over the full forecast horizon (mean field and producer-well error of 0.55 and 0.77~K, respectively), providing reliable estimates of the time of thermal breakthrough, which outperforms U-Net, Fourier neural operator, and MeshGraphNets baselines. The surrogate produces a full 100-year forecast in roughly a tenth of a second, $260\times$ faster than the full-physics simulation in batched ensemble mode, highlighting its potential for ensemble-based uncertainty analysis, optimization, and efficient geothermal reservoir management.