Author: Hemant Kumar
Regional Technical Sales Manager
Author: Azucena Gomez
Global Business Manager
The following article is the second in a 4-part blog series, which explains the four success factors that can improve the value and impact of AI/ML initiatives in the E&P industry.
The first article in the series is Enabling Artificial Intelligence and Machine Learning Through Talent Transformation.
As the adoption of Artificial Intelligence (AI) expands across the business sectors, so does the public debate around its limitations and vulnerability to malfeasance and human biases. The latest development is the eyebrow-raising concerns, fanned by extensive media coverage, about AI’s fallibility in the wake of the latter producing controversially questionable results with seemingly native tasks such as language understanding or facial recognition. This has added to a growing call for regulation and governance to drive responsible usage of the technology. AI suddenly finds itself haunted by the trust issues.
The E&P industry may be removed from the swirling public debate on AI, but trust has always had a certain influence on the industry’s AI initiatives. In our extensive interactions with the E&P professionals, we are told that they have difficulty trusting AI-based solutions introduced by their companies. Out of several factors, we have identified the four major reasons that generate the deficit of trust in AI.
Four Reasons for the AI Trust Deficit in the E&P Industry
Fig: Four reasons that cause trust deficit in AI in the E&P industry
How Halliburton Helps E&P Companies Overcome AI Trust Deficit
Halliburton, with experience of over 40 AI projects globally, understands the unique challenges of implementing AI in the E&P industry. The following diagram illustrates our key strengths in executing AI-led transformation initiatives of our customers.
Our approach leans on the three important differentiators.
Halliburton’s Data Scientists Understand E&P Domain and Challenging Datasets
Several E&P companies we have spoken to cite lack of domain knowledge as a reason for limited success with their AI projects. The traditional approach of pairing non-domain data scientists with SMEs has begun to lose favor in the E&P industry.
By virtue of their knowledge and experience, SMEs face significant demand on their time from various quarters in their organizations. It’s rarely a good (or fair) use of their time to handhold data scientists and spend time explaining the meaning of data, basics of workflows, and how to tell meaningful correlations from the obvious ones.
Our data scientists hit the ground running. They are well versed in both the science of building AI algorithms and the arcana of petroleum engineering and geosciences. Supported by an extended global network of industry thought leaders and R&D teams, our data scientists bring new ideas and insights from their experience with similar projects, thus helping uncover potential blind spots in data collection or defining success criteria.
Fig: Halliburton Data Scientists bring unique blend of skills and experience
We Bring Specialized Tools to Improve Data Quality and Enrich It
A lot of E&P data doesn’t come from the direct measurements. There is always some human touch involved which “contaminates” the datasets. Well logs or seismic data, for example, are interpreted to derive the properties of subsurface. Those interpretations may vary from one expert to another, depending on their level of expertise or other circumstances such as quality of the measuring equipment, conditions of the wellbore or mud infiltration. The same goes for well test data or fluid samples.
Our data scientists have the domain expertise to understand such peculiarities of E&P data. Using our domain knowledge, specialized software and proprietary dataset we can spot data inconsistencies issues and improve them. For example, our DecisionSpace® Data Quality platform is a widely adopted data platform used for faster identification of gaps in the large data sets. With our Seismic Engine, a DecisionSpace 365 application, we can improve the quality of fault exactions while applying AI to automate fault interpretation. Similarly, our proprietary Neftex® Predictions solution provides a rich account of basins worldwide that we use to enrich a client’s existing data sets.
Fig: Halliburton’s approach to dealing with challenging E&P datasets
Trust First Modeling Approach
Our study of over 100 SPE papers on artificial intelligence or machine learning projects in the E&P industry reveals a vast variety of model building techniques, each emphasizing their own advantages. What we can safely conclude is that there is no real best practice for choosing specific techniques to build AI models. It is a function of the nature of the field, prevailing workflows, available data and end users’ expectations about the explainability of the model. The last is an underappreciated but a significant determinant of acceptability of AI. The more complex the use case is, the more end users want to be able to explain the inner workings of the models in order to feel confident in them.
What complicates explainability is the non-linearity of asset behavior typically seen in oilfield operations whether it be fracking a well, injecting water into the reservoir, or producing wells through chokes. Simply put, non-linearity is the effect where increasing an input parameter doesn’t lead to a proportionate increase in a specific output variable. Non-linearity can also render purely data-driven models ineffective as soon as the real-world conditions stray outside the training data sets. For instance, a model trained to predict production rates will not predict water cut if the wells have never produced water before.
Halliburton tackles these challenges with a trust-first modeling approach.
Collaboration is at the heart of this approach. Our data scientists work with the client’s SMEs to develop the causal maps to discover and capture uncertainties and non-linearity in the current process. Those maps, along with designs of experiments, drives the selection of the most appropriate modeling technique meeting the following criteria:
Throughout our collaboration with end users, we ensure they have complete visibility into the working of the models and the assumptions and constraints that underpin them. The outcome of the exercise can be one single model, or a series of sub-models linked to each other. The user-friendly workflows and visualization tools provide end users with the ability to analyze the models and enhance them with the help of new data, when it’s available.
Fig: An illustration of Halliburton’s trust-first modeling approach for a subsurface property prediction use case
There is no shortcut to addressing the problem of the AI trust deficit. Good data, the right skills, leadership support, and healthy collaboration between data scientists and end users are all important ingredients to addressing the deficit. Getting them right takes sustained efforts and discipline, and E&P companies must commit to them in order to improve the adoption of AI. No two oil fields are the same and each E&P company will go through their own unique journey with AI to figure out what works best for them. Choosing the right partners, however, can accelerate the learning curve and ease the organization growing pains in building trust with AI.
Click here to learn more about what we’re doing in Talent Transformation in the oil and gas industry.