Prediction of Mineralogical Composition in Heterogeneous Unconventional Reservoirs: Comparisons Between Data-Driven and Chemistry-Based Models

Abstract

Abstract Prediction of mineralogical compositions along multi-fractured horizontal wells (MFHWs) using indirect methods, for the purpose of characterizing lithological and rock brittleness heterogeneity, is appealing due to the challenges associated with direct mineralogical evaluation. This study aims to 1) develop predictive machine learning models for indirect estimation of mineralogical compositions from elemental compositions, 2) compare mineralogical compositions obtained from data-driven and chemistry-based approaches, and 3) provide practical recommendations for fine-tuning and training of data-driven models. Leveraging recent advances in deep learning, an attention-based gated recurrent unit (AttnGRU) with a "feature extractor-post processor" architecture was developed for predicting compositions of ten primary minerals based on elemental data. For comparison, classic regression-based and ensemble learning models including support vector regression (SVR), random forest (RF), and a feedforward neuron network (FFNN) were utilized. Data-driven models were trained and tested using XRD data measured on 217 samples from the Montney Formation, and the outcomes were compared to those derived from stoichiometric material balance equations (a previously-developed chemistry-based model) to evaluate the effectiveness and capabilities of different predictive approaches. The data-driven models consistently outperformed the chemistry-based method with significantly lower mean absolute error (MAE) and higher R2. The predictive performance order was FFNN ≥ AttnGRU > RF > SVR >> chemistry-based model, with MAE = 1.05, 1.09, 1.24, 1.35, and 2.46 wt.%, respectively. Importantly, FFNN, AttnGRU and RF offered more accurate predictions of chlorite and illite, which are known to negatively affect reservoir quality. This indicates the superior performance of the three models for reservoir characterization applications. Furthermore, AttnGRU exhibited greater robustness than the other two models, with less sensitivity to overfitting issues. Data-driven models displayed different levels of performance when decreasing training dataset size. It is recommended that, in order to achieve reasonable predictions for the studied reservoir with data-driven approaches, more than 50 training samples be used. It is further observed that data-driven models exhibited limited predictive capability (MAEs ranging from 3.02-3.45 wt.%) when applied to a synthetic "global dataset" comprised of samples from various formations. Through the comparison of multiple independent datasets (XRF-derived chemistry-based, XRF-derived data-driven, XRD) collected on identical samples, this work highlights the strengths, limitations, and capabilities of different machine learning techniques for along-well estimation of mineralogical composition to assist with reservoir characterization.