Accuracy and Interpretability – Making data science work for you

As oil and gas (O&G) companies continue towards investing in digital technologies to drive business value, it's clear that data science will play a big part in helping drive that success. In a recent Gartner 2021 CIO Survey, about 50% of oil and gas companies say they plan to increase their investments in analytics, Artificial Intelligence (AI), automation, Internet of Things and cloud technology. This is not surprising, as the tool becomes more intuitive and computing power rises.

The O&G industry has long been a pioneer in computing use. From building some of the world's first supercomputers to processing terabytes of seismic data to complex reservoir simulation, engineers and geo-professionals have often been on the leading edge of using data and models to unlock the earth's secrets and improve operational efficiency.

However, todays' journey to AI is still a challenge for most O&G companies. It's well known that adopting this technology in the industry is slow - but it doesn't have to be. An area in Upstream E&P that can reap the benefits of this technology is petroleum engineering as they are continually finding new ways to optimize production and costs. Today, they can combine their deep domain expertise with data-driven analytics to enhance and develop better asset strategies. Within 5-10 years, IHS Markit believes many AI/Machine Learning (ML) applications will move from being the purview of a specialized team of dedicated data scientists to become a new set of tools used regularly by a wide range of engineers across the company.

Upstream E&P: What Engineers Can Expect from an AI Solution

Under current O&G market conditions, engineers face many challenges that require them to find more efficient ways to solve complex subsurface problems. Part of that responsibility involves analyzing data from disparate streams and building models that feed into their drilling, completions and production strategies. So, the importance of model accuracy is vital to ensure confidence in results and in decision-making.

Modern advanced data analytics solutions built for Upstream E&P provide the advantage of AI without requiring specialized data scientists, data managers and domain experts writing complex code. The best solutions analyze vast amounts of data rapidly and provide interpretability to help users better understand the model results. Visual displays allow for efficient interrogation and analyses. Even more sophisticated solutions provide workflow templates that incorporate ML algorithms for addressing known E&P challenges.

Built with existing well and production basin data, these analytics platforms are flexible enough that users such as petroleum engineers can incorporate proprietary datasets, validate results and rerun their predictive models as more data and knowledge accumulates.

Analytics Explorer dashboard analyses of multiple data sources to be used in a predictive model. Results show the order of importance of variables and which ones should be used to create the predictive model.

Unboxing the Black Box

AI predictive models have been called Black Boxes: data is put into the model; a user hits Run and out spits a number. It's easy to understand why engineers are often skeptical of this approach. They learned long ago that models can be made to output nonsensical results. The IHS Markit models of today are designed to be very different. Users can - and should leverage their expertise to provide reality checks, verify the data correlations and compare outputs to ground truth numbers. Detailed error analysis is available for each model, allowing users to review the error associated with the inputs and final results. Users can identify which inputs are causing high error, address them, and rerun the model.

Importantly, the IHS Markit predictive tools do more than predict future wells. They also quantify how much different well attributes-for example, well path tortuosity, location or proppant volume-are contributing to the final model results. Users can choose which data goes into the model, evaluate the importance of each attribute, and remove the data that does not make a meaningful contribution. Removing noise and unimportant data strengthens the model and provides better results. Even more importantly, the users can test hypotheses to determine whether potentially important factors are impactful or not. They can be open to surprises and serendipity.

Accuracy and Interpretability

Perhaps most importantly, the results of IHS Markit's modern predictive models provide that elusive combination of results that are both accurate and easy to interpret, even for non-data scientists. For example, starting with a crossplot of actual 12-month production compared to the results predicted from the model, users can select any data point and see quantitative information about the impact on production of each and every attribute in the model. The crossplot also shows the strength of the correlation, so users can measure the model's deviation from the truth data.

A crossplot of actual 12-month production data (x-axis) and predicted machine learning model (y-axis) results. The bar chart shows the contribution of the input variables to production for a selected well.

Reach Better Outcomes with Analytics Explorer

IHS Markit is making advanced data analytics technology easier to adopt and deploy, giving engineers the ability to utilize data science methodologies along with their domain expertise to solve complex subsurface problems.

Analytics Explorer is an advanced data analytics solution from IHS Markit that makes data science accessible to everyone. Developed for Upstream E&P, Analytics Explorer incorporates advanced data science methodologies in guided and automated workflows that incorporate interpretability methods to help engineers better understand their models with confidence.

With customizable workflows like predictive modeling, Analytics Explorer can be used for a broad range of applications, including:

• Predicting the performance of wells before drilling them
• Quantifying the impact of specific parameters on a well's performance
• Understanding optimum well design
• Understanding the impact of location and completion quality
• Identifying re-frac candidates
• Benchmarking performance among operators

For more information, download the ebook for a predictive modeling workflow using Analytics Explorer.

Posted 28 March 2022 by Camilo Rodriguez, Director, Analytics and Data Science,S&P Global Energy and

Raoul LeBlanc, Vice President, Energy, S&P Global Commodity Insights and

Toby Burrough, Senior Technical Advisor, S&P Global Commodity Insights

This article was published by S&P Global Commodity Insights and not by S&P Global Ratings, which is a separately managed division of S&P Global.