Novi Technical Advisor Ted Cross will be presenting at the SPE Dallas Section Study Group, covering the basics of machine learning: when to use machine learning, data preparation, measuring performance, how to explain model results, and leveraging the model for optimization scenarios. The presentation will then review applications of these concepts in the Midland Basin using a public-data model, looking at completions sensitivities, geology-spacing interactions, and asset development optimization case studies.

What is the minimum amount of geological data required to build a predictive model? Traditional basin modeling approaches utilize expensive datasets and significant time expenditure to understand and predict fluid distribution around the basin. Machine learning offers an empirical method to generate similar predictions. In contrast to the forward-modeling method of traditional basin modeling, machine learning looks for implicit relationships between training and target variables. In this study, we investigate whether a machine learning model fed only geologic tops as input geological data can accurately predict well performance in the Bakken-Three Forks play of the Williston Basin.

Not all rocks should receive the same completion. It’s a simple concept, but it can be tough to back up the theory with statistical evidence. Machine learning provides a powerful tool to analyze the interactions between completions design choices and local geology. In this talk, we analyze interactions between proppant, fluid loading, stage spacing, and geology. We then show that per-well economics can be improved by over $1MM NPV by taking a tailored approach. SHAP values (like in the image above) show what the model is thinking: what exact combinations of geology & spacing are driving these interactions.

Understanding the impact of interwell spacing is tough enough before introducing the impact of parent-child depletion. When you add in the fact that both have a strong correlation with geology—especially in the Bakken, where much of the early producers and early downspacing tests were in the best rock–well, then things can really get hairy. In this talk, we investigate a variety of ways of measuring parent-child and spacing impact, including days online, number of parents within a radius, parent production, and distance to neighbors.

Applying a single scaling factor over the life of the well for a completions variable will give you an inaccurate production profile — and inaccurate economics! Quantifying the time-variant impact of completions decisions is tough for traditional methods. However, training a machine learning model to predict a time series of production, combined with Shapley values, provides a powerful tool to investigate when completions are most impactful.

Like it or not, most companies don’t spend nearly as much time analyzing water production as oil (and who can blame them!?!). Fortunately, machine learning techniques developed for forecasting oil production can be adapted for water production — it’s part of our standard model delivery.

Canadian operators have an information problem: in the public data, “gas” wells don’t have to report their liquids content. Novi has developed a solution to estimate CGR that’s twice as accurate as the traditional methods.

Since Novi was founded, customers, partners, and prospects have been asking us if machine learning could be used to extract a subsurface model directly from logs. The answer is yes.

In this paper, we analyze the changing impact of interwell spacing through time on oil production in the Delaware and Midland Basins.

How do parent wells impact the GOR of their children?

With the right dataset and right machine learning model, you can get high-resolution answers on fine-tuning your target selection — potentially increasing production by over 10% at zero cost.

How can you properly benchmark operators — to identify overperformers to learn from, or underperformers to acquire?