Machine Learning Optimizes CO₂-Enhanced Oil Recovery

Summary

A high school-level preprint applies a Random Forest classifier to 1,000 simulated CO₂-EOR scenarios generated from the SPE10 Model 2 benchmark. The goal is to identify the injection and pressure conditions that most strongly impact oil production and profit.

Why it matters

CO₂-EOR is a significant lever for both enhanced oil recovery and carbon capture. Consequently, decision-support tools developed around injection rate and pressure management hold real operational value. However, the contribution here is primarily pedagogical. The study demonstrates that a standard machine learning pipeline can be integrated with a known reservoir benchmark. While this is useful for training the next generation of petroleum data scientists, it does not advance the state of the art in the field.

Highlights

• A Random Forest trained on only 1,000 SPE10 scenarios to rank feature importance for oil output and profit
• A physics-based 'digital twin' based on the standard SPE10 Model 2 benchmark (10×10×5 grid)
• The author is a high school student; data and code are not yet publicly available

Critical take

Labeling a Random Forest study based on a single benchmark as a 'digital twin' stretches the term. Furthermore, there is no baseline comparison (e.g., gradient-boosted trees, Bayesian optimization, or a simple response surface model). This makes it difficult to assess whether machine learning genuinely adds value compared to the underlying physics simulation. With only 1,000 scenarios, a lack of uncertainty quantification, and no external validation dataset, the practical transferability to real, heterogeneous fields remains unproven.