URTeC 2026 - Technical Paper

Making an Impact: Machine-Learning-Ready Metrics for Ultra-Tight Infill Development

How can we better model ultra-tight infill development? This URTeC 2026 paper explores machine-learning-ready metrics that improve spacing analysis and production forecasting.

Making an Impact: Machine-Learning-Ready Metrics for Ultra-Tight Infill Development

Line graph showing % degradation from 3 well-per-section for 3-year oil cumulative versus average distance to 2 closest neighbors, incorporating infill development spacing metrics, with an inset scatter plot of well placements.
Theme 08

Connecting Forecasting with Reservoir Management Decisions

Topic

Making an Impact: Machine-Learning-Ready Metrics for Ultra-Tight Infill Development

Authors

B. Davis*1, A. Qualls2, K. Sathaye1 and G. A. Quintero1 (1. Novi Labs; 2. Devon Energy)

Abstract

Objectives/Scope: In this paper, we examine different approaches to assessing infill spacing impacts on machine learning model outputs. This work will focus on Devon-operated assets and will assess various methods of describing spacing impacts for the purposes of predicting future well performance. Our goal is to examine the benefits of a machine learning approach to infill well forecasting. In many cases, machine learning approaches handle spacing impacts properly for typically spaced wells. However, naïve spacing metrics struggle with spacing designs that are much tighter than the basin average.

Methods/Procedures/Process: As a source of actuals, we used an anonymized dataset of wells from Devon operated pads in the Delaware Basin, covering a variety of infill spacing scenarios. We then trained causal inference, random forest models using different feature sets. These models were used to predict production for a set of Devon operated wells which served as a blind holdout test set. Performance was measured against actual production for the hold out set. Our baseline method for this work was the pre-drill planning type-curves used for these wells. Our evaluation features were traditional point-to-point distances (tangent distance), distance weighted features (impact factors), area based features (Voronoi tiling), and combinations of these different approaches (Voronoi + Impact, Voronoi + Tangent, etc).

Results/Observations/Conclusions: There is a significant improvement in accuracy and spacing sensitivity when combining causal inference, random forest models and distance weighted spacing features. While using the features alone (only Voronoi or only impact factors) yield good results, we find the best performance is a combination of these features. In our test set, type curve forecasting methods showed 20% median absolute percent error rates against actuals for 1-year cumulative production. In contrast, machine learning based forecasts showed a 15% error rate when judged on the same metric. SHAP explanation value analysis shows that while the more digestible untransformed features showed directionally correct trends, only the transformed features captured the expected non-linear spacing performance degradation.

Applications/Significance/Novelty: Error metrics for machine learning models are one facet of tool evaluation. However, they don’t present the complete picture. Different feature sets can have similar errors while also exhibiting different sensitivity to spacing changes. It can be difficult for engineers to consider all the interacting physical processes such as geology, completions, spacing, and depletion. Machine learning approaches, especially methods that are aligned with good engineering judgement, are an integral tool in evaluating development planning and optimization.

Complete the form and we will email the paper straight to your inbox.

[URTeC 2026] - Papers Requests

First name(Required)

Related Resources

URTeC 2026 – Technical Paper Designing Depletion Features for Parent–Child Modeling: Comparing Time, Distance, and Volume–Distance …

URTeC 2026 – Technical Paper Well Spacing Impacts Water Production in The Permian Basin How does …

URTeC 2026 – Technical Paper Making an Impact: Machine-Learning-Ready Metrics for Ultra-Tight Infill Development How can …

Meet Novi Labs' knowledgeable experts.
Invite to Speak

Please provide us some information about your event, and we will get back to you as soon as we can

Expert form

This field is for validation purposes and should be left unchanged.
News, trends and data for the US upstream industry
Receive latest updates on the energy industry into your inbox.

Sign up today for free and get a weekly email with News and insights on important developments in the major tight oil & gas basins.

Subscribe form (popup)

"(Required)" indicates required fields

This field is for validation purposes and should be left unchanged.