Causality-driven feature representation for connectivity prediction

Area of Science:

• Petroleum Engineering
• Causal Inference
• Machine Learning

Background:

• Causal reasoning is crucial for understanding complex systems and decision-making.
• Oil field management requires identifying injector-producer well connections for optimization.
• Controlled experiments are infeasible, necessitating reliance on observed data.

Purpose of the Study:

• Develop a causality-inspired framework for robust injector-producer connectivity estimation using observed data.
• Address challenges like confounding factors, system response latency, and inter-well complexities.
• Leverage domain expertise for causal feature learning to improve accuracy.

Main Methods:

• Framed the problem using causal inference principles.
• Proposed a novel framework generating pairwise features driven by causal theory.
• Constructed independent pairwise feature representations to implicitly handle confounders.
• Utilized limited context data for training machine learning models to estimate connectivity probability.

Main Results:

• Validated the methodology on synthetic and semi-synthetic datasets.
• Applied the framework to Brazilian Pre-Salt oil fields using real-world data.
• Demonstrated effective identification of injector-producer connectivity with rapid training times.

Conclusions:

• The proposed method offers a scalable and interpretable approach for connectivity estimation in complex dynamic oil field systems.
• This work represents a systematic formulation using causal reasoning to address confounders and discover inter-well connections.
• The framework enhances prediction accuracy and decision-making in oil field operations.