Transforming Academic Research: Solving Previously Complex Oil and Gas Problems Using Machine Learning

Digital transformations are slated to transform the industry by reducing expenditures, improving operations, and providing a granular view of workflows enabling more effective decision-making. In the heart of all these digitization efforts in our industry lies machine learning.

Machine learning enables us to build complex models on the data collected, leading to better decisions. In the simplest terms, it is a form of artificial intelligence (AI) which is designed to learn on its own or become better as it is fed more data. These algorithms have the potential to revolutionize our workflow in the future when the applicability of AI increases. However, some machine learning algorithms can be easily incorporated into our workflow today to provide added value.

Effective application of machine learning to solve real-world problems in our industry is becoming a reality thanks to the availability of computational power, advances in algorithms, effective collection of data, complexity of problems, and digitalization efforts. 

This article highlights interesting applications of machine learning in the oil and gas industry in drilling, formation evaluation, and reservoir engineering. Each project uses a data-driven model to solve a previously complex problem using machine learning to augment an existing solution.

Drilling: Coupling Downhole Parameters

Chiranth Hegde is a data scientist at Shell where he works on incorporating machine learning and artificial intelligence to improve drilling performance. He earned his MS in 2016 and PhD in 2018 from the University of Texas at Austin where he developed machine learning algorithms to optimize drilling. He completed his undergraduate studies in mining engineering from the National Institute of Technology Karnataka, Surathkal, India. He is a member of the TWA Editorial Committee.

Chiranth Hegde recently obtained his PhD from the University of Texas at Austin working with Ken Gray in the Wider Windows research consortium. He is using machine learning to develop an end-to-end coupled drilling optimization advisor. Most optimization strategies rarely include the effect of vibrations, which is the biggest rate of penetration (ROP) inhibitor. As such, most optimization studies have attempted to optimize only one drilling metric – ROP, mechanical specific energy (MSE), or vibrations – independent of the other, despite their interaction downhole. This does not represent bottomhole conditions accurately. Developing a coupled model can be extremely challenging given the complex interaction of different parameters downhole.

Hegde’s research introduces a method of coupling several downhole parameters using machine learning algorithms. Models for ROP, torque on bit (TOB), MSE, and stick-slip are built using a data-driven modeling approach using the Random Forests algorithm. The models are coupled by building them conjointly during training. The coupled model was evaluated to optimize ROP and MSE on validation data. ROP increased by 31% and MSE decreased by 49% on average with the use of this optimization model. The model can be used as a real-time drilling advisory system and provide heat maps which can help guide the driller.

Fig. 1—Heat maps for ROP, TOB, and MSE in Lodgepole Limestone formation as determined by the coupled data-driven model. The true optimal values of each metric have been plotted as a cyan star. The optimum settings can be chosen based on the metric of interest.

Formation Evaluation:  Improving Subsurface Characterization

Hao Li is a PhD candidate in the Mewbourne School of Petroleum and Geological Engineering at the University of Oklahoma. His research includes well logging, machine learning, deep learning, petrophysics, and unconventional enhanced oil recovery. He holds an MS degree in petroleum engineering from China University of Petroleum-Beijing.

Hao Li is trying to improve subsurface characterization using deep learning. Hao is a PhD student working with Siddarth Mishra at the University of Oklahoma. NMR logs provide exceptional insight into fluid composition and pore size distribution of a formation. However, it is expensive to acquire. Li’s research aims to use deep learning to create a synthetic NMR log using deep learning.

Deep learning algorithms are generally used for prediction of a point value (regression) or classification, which involves the training of a discriminative model. However, they can be modified to produce samples from a distribution using a generative model by modeling the joint distribution.

A popular successful approach is the use of variational autoencoders (VAE) as shown in Figure 2. The basic idea is to first train a decoder or a generator, which generalizes the features in the NMR training set. The trained decoder/generator recognizes the basic “genre” of NMR T2 in a reservoir. After the decoder/generator learns the NMR T2 features, a simple neural network is used to associate a different kind of NMR T2 with a different combination of conventional logs. Deep learning algorithms may enable us to see what it has learned instead of being a “black box.”


Fig. 2—Two-step training process of variational autoencoders.


Reservoir: Finding a Better Solution for Well Placement

Azor Nwachukwu is a PhD student in petroleum engineering at the University of Texas at Austin. He received a BS degree in chemical engineering from the University of Maryland Baltimore County in 2013 and an MS degree in petroleum engineering from the University of Texas at Austin in 2015. His research interests are machine learning, geostatistics, and reservoir characterization.

Azor Nwachukwu, a PhD student at the University of Texas at Austin, is working with Michael Pyrcz to determine a better solution for well placement. The conventional solution is to use a reservoir simulator as a function evaluator which can be extremely computationally expensive. While a data-driven model can be used in place of the reservoir simulator, the geological complexity of reservoirs can make this an infeasible solution. Models are also dependent on geological complexity; a predictor trained using a certain geologic model would fail to make accurate predictions when tested with a different equiprobable geologic realization.

Nwachukwu uses a machine learning-based proxy to solve the computational inefficiency of reservoir simulators. However, the model is augmented with connectivity networks to incorporate physics into an otherwise purely data-driven approach. The connectivity network contains data related to pairwise well-to-well connectivities, spacing, and overall coverage.

These networks introduce the concept of geological realizations into the model. Prediction models were trained using the gradient boosting algorithms. The model when tested against diverse scenarios and case studies shows promising results.


Fig. 3—Log10 permeability of the 10-model ensemble showing producer location.


Fig.4—True vs. predicted objective function (NPV) using 200 training observations (reservoir simulations). Left: With connectivity network. Right: Without connectivity network.



Based on the examples shown in this article, it is evident that machine learning can help solve practical problems in our industry. Currently, the research in academia is focusing on the utilization of machine learning to solve previously unsolvable or challenging problems. Future applications of AI such as automatic interpretation of well logs, fully automated rigs, offshore rig maintenance identification using drones and computer vision, and fully automated reservoir management are not very far off. Combining these efforts with the industry’s push for digitalization, it looks to be an exciting time ahead.


Stay Connected

Don't miss our latest content, delivered to your inbox monthly. Sign up for the TWA newsletter.  If you are not logged in, you will receive a confirmation email that you will need to click on to confirm you want to receive the newsletter.