Title: Regional to Local Machine-Learning Analysis for Unconventional Formation Reserve Estimation: Eagle Ford Case Study

Authors: Peidong Zhao, Rencheng Dong, Yu Liang

SPE/OnePetro ID: SPE-201351-MS

keywords: Eagle Ford Shale, Estimated Ultimate Recovery, Big Data, Machine Learning, Random Forest, K-Nearest Neighbor

Unconventional tight reservoirs currently make up more than 60% of domestic oil and gas production in the United States. However, developing unconventional formations requires intensive drilling and completion campaigns to maintain steady production of a field. Therefore, the prediction of estimated ultimate recovery, which measures the producible reserve from a well, is demanding, particularly as operators becomes more rational under the current volatile market conditions. Despite unconventional reservoirs being considered a resource play with low geological risks, their economic appraisal is challenged by unknown stimulation outcomes and intricate producing mechanisms. Therefore, this work aimed to leverage machine-learning techniques with big data to analyze the multivariant relationship of geological and engineering parameters with unconventional reservoir production and to improve the prediction of estimated ultimate recovery in unconventional formations.

The objective of this work was to improve the prediction EUR in eagle ford formation. A multiscale approach was adopted in this work. The large-scale trend model aimed to delineate sweet spots, which grants direct guidance for acreage acquisition and development across the basin. The small-scale predicting model aimed to collect key parameters in anticipating high production at a specific area and insights for production optimization in the field.

In this case study, a multiscale machine-learning workflow was deliberated and applied to a big data set from the Eagle Ford shale. First, quality control and feature selection were performed on the big data set (4,067 wells, 30+ features), in which statistical correlation and domain knowledge were both considered. Then, taking advantage of the big data set densely covering the Eagle Ford and the combination of KNN and bagging techniques, the region-scale trend model aimed to obtain preliminary inferences on sweet spots and future drilling locations. With spatial similarity demonstrated, a local-scale model regression leveraging additional geological and engineering features was developed to finalize EUR prediction. Along with our previous approach of separating oil EUR and gas EUR, the classification of well location further improved the accuracy of EUR prediction. At the end, the data set was reviewed to validate the results of the physics-guided data-driven model. Through the workflow, every step revealed key information about Eagle Ford EUR, eventually leading to a far-reaching comprehension of the data set. Therefore, this progressive learning strategy was proved to be practical for unconventional formation reserve estimation.

The application of the proposed workflow on the Eagle Ford shale demonstrates a progressive building of the machine-learning model. The quality control of data allows global inspection of the data set and, more importantly, confirms the statistical distribution of training data. This study emphasizes the philosophy of multiscale data analytics. The large-scale model portraying sweet spots using location variables grants direct guidance for acreage acquisition and development across the basin; the small-scale model trained with reduced dimensionality generates quantitative prediction of oil and gas estimated ultimate recoveries for an area of interest. Compared with our previous work using higher dimensionality and extensive spatial interest, this progressive learning maintains similar explained variance in out-bag model check, but grants 26% and 52% reductions in mean square error for predicting oil and gas estimated ultimate recoveries in the Eagle Ford formation. In the end, prediction validation is performed by revisiting the data set. Overall, the proposed workflow demonstrates successful application in the Eagle Ford formation such that it can be directly implemented for other unconventional resource plays.

The contributors of this repository declare no competing interests. This repository is for educational purposed only. For additional information request, please contact

• Yu (Reservoir+Production+Completion, [email protected]) for Dataset and ML regression
• Peidong (Geomechanics+Drilling+Completion, [email protected]) for Trend Model and ML regression
• Rencheng (Reservoir+Production, [email protected]) for Geostatistical Model and Trend Model

Dataset or sensitive information is unpublished for confidentiality.

The training data for each model are selected from total of 4,067 wells based on two criteria: credential source and statistical distribution.

1. Trend Model (KNN-bagging Algorithm)

   • 1st screening: select wells from operator with 30+ wells in the basin
   • 2nd screening: remove wells with super-high EUR ( Oil > 4.5e5 bbl, Gas > 4.5e6 boe)
   • Training Data: 3,000+ wells across Eagle Ford
   • Predictor Features: relative location
   • Respondse Features: Oil EUR, Gas EUR

2. Predicting Model (Random Forest Regression)

   • Multicolinearity mitigation: pair-wise Pearson correlation coefficient
   • Domain Knowledge: Geology, Geomechanics, Reservoir Engineering, etc.
   • Training: 200+ wells in East Texas
   • Predicting Features:
     • well location, well depth, lateral length, horizontal azimuth, and proppant intensity (lb/ft)
     • oil type, TOC, VRE, hydrocarbon pore volume, compressive strength, Young's modulus, pressure gradient, Upper Eagle Ford thickness, Lower Eagle Ford thickness, , density of producing fluid
   • Respondse Features: Oil EUR, Gas EUR

In this study, we present a multiscale machine-learning workflow to predict EUR in unconventional formations. The latest approach reduces predicting error by more than 50% compared with our previous work on the Eagle Ford shale. The workflow consists of four parts: data collection, data preprocessing and quality control, trend modeling, and regression modeling.

The deliberated quality control of data and feature selection underwrites the predicting accuracy of machine learning. Feature selection was performed iteratively using machine-learning results from previous work while being guided by domain knowledge. The reduced dimensionality proved beneficial in the subsequent model building.

The trend model adopted a KNN-bagging algorithm. With well location being the only training feature, the EUR trend was mapped across the formation. Results from the trend model delineated multiple sweet spots within the Eagle Ford region, implying locations for future drilling. The regional-scale trend model also indicated two distinct trends in South Texas and East Texas. With that, the East Texas area was cropped for the local-scale regression model. Compared with simple kriging results, the KNN-bagging method shows the advantage of simplicity and adaptability for large-scale data sets with complicated spatial variations.

The regression model selected the random forest algorithm to predict oil and gas EURs in the East Texas Eagle Ford formation. The random forest regression was performed under rigorous cross-validation and hyperparameter tuning. With reduced training features and spatial interest, the explained variance was the same for gas EUR and 61% decline for oil EUR, but the mean square error of testing achieved striking reductions of 52% for gas EUR and 26% for oil EUR. The results of feature ranking indicate influence from the randomness of training and test split and random forest algorithm, so we implemented an ensemble approach to resolve the issue.

In the end, we revisited the data set to validate the machine-learning results. The high-production wells are clustered within a specific range of geological features. However, regarding the dispersed distribution of EUR at a promising feature value, we suggest developing probability-based models for future work to assess subsurface uncertainty. The review of engineering features indicates the needs of the latest data, which would expand the learning scope of the computer algorithm. To further extend this workflow, we also specify to include operational cost to perform economic optimization of the well using the regression results.

Overall, with the assistance of the progressive learning and guidance of domain knowledge, this work shows boosted capability in predicting Eagle Ford EUR as compared to our previous work. The workflow presented in this work can be directly adopted to other unconventional plays.

The authors wish to thank Prof. Michael Pyrcz and Honggeun Jo at the University of Texas at Austin for their helpful discussions during the courses of geostatistics and subsurface machine learning. Our program has applied open-source packages: scikit-learn (https://pypi.org/project/scikit-learn) and GeostatsPy (https://pypi.org/project/geostatspy).