Oil and Gas Exploration
Oil and gas exploration is fundamentally a decision-making problem under extreme uncertainty. Subsurface structure is only partially observed through noisy measurements such as seismic responses, well logs, pressure data, geochemistry, and sparse drilling outcomes.
While physical laws govern wave propagation and fluid flow, there is no closed-form equation that maps observations to exploration outcomes and geological systems are strongly non-stationary, path-dependent, and heterogeneous across space and time.
Recent AI/ML approaches have largely focused on applying deep learning models to seismic images or engineered features. While useful for perception and interpretation tasks, such models tend to smooth structure, obscure uncertainty, overwrite historical evidence, and extrapolate beyond what the data can support. The result is a persistent gap between what is known empirically and how that knowledge is operationalized in exploration decisions.
Our Approach
We treat subsurface exploration not as a prediction problem, but as an information decoding problem. The subsurface can be viewed as a noisy information channel where geological structure is encoded into physical signals by wave propagation and sampling geometry, corrupted by noise and incomplete coverage, and observed through indirect measurements. The objective is not to “learn geology,” but to decode repeated structural states from partial, noisy observations and aggregate empirical outcomes associated with those states.
Our system applies an information-theoretic, nonparametric framework built around three principles:
1. Explicit encoding of observed structure
Subsurface observations (structural context, seismic attributes, depositional indicators, depth, pressure regime, maturity proxies) are transformed into symbolic representations. This encoding is intentional and controlled, preserving magnitude and structural distinctions rather than smoothing them away.
2. Empirical memory of outcomes
Each encoded state is stored alongside observed outcomes from historical wells or prospects. No global model is fit, and no parameters are optimized. Knowledge accumulates monotonically as evidence. Each well drilled, each core analyzed, each seismic survey processed adds to the empirical base without degrading or overwriting what came before. The system’s reliability improves strictly with data, unlike models that can degrade when retrained on new data.
3. Decoding at inference time
New prospects are decoded by comparing their encoded state to previously observed states using information-theoretic distance metrics. Similar historical states are identified, and empirical outcome distributions are aggregated with explicit confidence measures based on evidence density.
This approach formalizes analog reasoning in a way that is quantitative, reproducible, and probabilistic, without assuming continuity, stationarity, or a learnable global function.
Example: Analog-Based Prospect Evaluation
A prospect in the Permian Basin is encoded using only information available prior to drilling:
• Structural position: Anticlinal closure at ~8,500 ft
• Seismic response: Class II AVO behavior
• Depositional environment: Carbonate platform margin
• Thermal maturity: Vitrinite reflectance ≈ 0.8%
The encoded prospect is decoded against the historical archive of subsurface states. The system identifies multiple prior wells with structurally similar encoded configurations across multiple basins.
Empirical outcomes for these analogs show:
• Commercial success: 68% based on analogs
• Estimated ultimate recovery (EUR): mean 425 MBoe (range 50–1,200 MBoe)
• Confidence: High (based on 47 independent analogs)
Each analog is fully inspectable. The geoscientist can review why each match occurred, examine the original context, and incorporate local geological knowledge before making a final decision.
Why This Differs from Conventional AI/ML in Oil & Gas
Most current AI approaches in exploration attempt to learn a mapping from data to outcomes. This implicitly assumes that subsurface behavior can be approximated by a smooth, stationary function; an assumption that does not hold in heterogeneous geological systems.
Our approach makes no such assumption.
• We do not train predictive models.
• We do not overwrite historical knowledge during retraining.
• We do not smooth across unobserved geology.
• We do not hallucinate confidence where evidence is sparse.
We preserve uncertainty explicitly and allow inference only where sufficient empirical evidence exists. When no meaningful analogs are present, the system refuses to extrapolate; a critical safeguard in high-stakes exploration where false confidence is often more costly than uncertainty.
Core Use Cases
• Prospect Analog Matching
Prospects are encoded using only information available at decision time and decoded against historical analogs. The system reports empirical success rates, outcome distributions, and confidence levels based on actual evidence density.
• Play and Fairway Analysis
Regional subsurface states are discretized and evaluated empirically, producing explainable fairway maps grounded in historical outcomes rather than smoothed probability surfaces.
• Well Planning and De-Risking
For proposed wells, the system identifies structurally similar historical wells and reports conditional risks and expected ranges, enabling transparent, evidence-based decision making.
A Unifying Framework
This oil and gas application is not a new model or domain-specific heuristic. It is the same information-theoretic framework we have validated in:
• Markets, where prices are treated as a noisy information channel revealing repeated state structure.
• Materials science, where atomic and molecular configurations are treated as discrete structural states with empirical properties.
Across domains, the method is the same: encode structure, accumulate evidence, and decode locally under uncertainty.
Why This Approach Is Timely
The oil and gas industry faces a unique convergence:
• Data volumes are exploding (high-res seismic, distributed sensors, basin-scale databases)
• Experienced geoscientists are retiring, taking analog knowledge with them
• AI/ML vendors are over-promising and under-delivering on exploration applications
• Capital discipline demands better risk quantification, not just point predictions
Our system addresses all four challenges:
• Scales to large datasets without loss of fidelity
• Captures and formalizes expert analog reasoning
• Avoids the pitfalls of overconfident AI/ML extrapolation
• Provides explicit uncertainty quantification for capital allocation
Implications
Exploration has always relied on analogs. What has been missing is a rigorous, information-theoretic way to encode those analogs, measure similarity under noise, and quantify uncertainty without imposing unjustified models.
Our system does not replace geoscientists or interpreters. It provides a formal memory and inference engine that makes analog reasoning explicit, probabilistic, and auditable; qualities that become increasingly important as data volumes grow and decisions become more capital intensive.