SPE DSEATS Datathon: Oil Presence Prediction & Physics-Informed ML

Context

Subsurface exploration requires navigating complex petrophysical measurements with sparse, noisy data. Purely statistical machine learning models often predict physically impossible subsurface conditions without domain constraints.

Challenge

Accurately predicting binary oil presence across complex geological formations while handling missing logs, rock-type-dependent properties, and maintaining physical validity under unseen test conditions.

Key Questions

• How do purely data-driven algorithms compare against physics-constrained models in subsurface characterization?
• What rock-physics rules (e.g., shale permeability limits and boundary clamping) provide the greatest regularization against overfitting?
• How can KNN imputation be constrained by trap types and rock lithology to prevent physical leakage?

Approach

• Engineered a dual submission pipeline: Part 1 utilizes a pure Gradient Boosting Classifier across 100% of the training features.
• Built a domain-aware physics-informed pipeline enforcing physical rules including shale permeability overrides and rock/trap-type KNN imputation.
• Applied physical boundary clamping and feature scaling prior to training a Support Vector Machine (SVM) model.
• Validated submission pipelines and output formats against competition benchmark specifications.

Outcomes

• Delivered high-confidence predictive submissions for both blind data-driven and physics-informed benchmarks.
• Demonstrated that incorporating geological domain rules significantly stabilizes model predictions in low-data regimes.
• Created a clean, modular Python codebase for repeatable feature preprocessing and test-set inference.

Takeaways

• Domain physics acts as an essential regularizer when training machine learning models on geological data.
• A dual-track strategy (pure ML vs. physics-informed) allows benchmarking raw model capacity against domain-consistent constraints.
• Data preparation in geoscience must respect lithological constraints rather than relying purely on generic imputers.