Geo Petro Litix AI – Physics Informed Machine Learning
Physics-guided Artificial Intelligence for reservoir characterization without hallucinations.
We train our models on hydraulic flow units (HFU) extracted from actual well diagenesis, not on statistical averages. We eliminate typical AI hallucinations by integrating physical laws into every prediction. And every result is fully explainable with TreeSHAP: there is no black box.
MAXIMIZING VALUE WITH LOWER CAPEX
Operating companies face growing pressure to reduce CAPEX and optimize asset performance. In this context, technical decisions must rely on better utilization of existing data to reduce uncertainty, lower costs, and increase the economic value of reservoirs.
Our solution applies Physics Informed Machine Learning to estimate petrophysical properties with high fidelity, avoiding the hallucinations presented by purely statistical models. Thus, each prediction respects the laws of reservoir physics.
CURRENT CHALLENGES FOR OPERATING COMPANIES
Operating companies must make high economic impact decisions with limited information. The quality of available data is a decisive factor in reducing uncertainty and improving outcomes.
Conventional AI can generate plausible but physically impossible answers (hallucinations). Our Physics Informed ML approach incorporates thermodynamic and flow constraints directly into the model's loss function, ensuring that every output is physically consistent. Furthermore, with TreeSHAP, we know exactly why the model made that decision.
PIML: CORRELATING THIN-SECTION, CORE, AND WELL LOG DATA
Our methodology integrates three historical data sources to train non-hallucinating models:
- Thin-section points: we extract coordinates from mineralogical point counts and link them to diagenesis.
- Core laboratory results: porosity, permeability, capillary pressure, and saturations.
- Well logs: GR, RHOB, NPHI, resistivity, among others.
With these data, we train a Physics Informed model that estimates petrophysical properties without losing physical coherence. The result: an updated static reservoir model, ready to re-evaluate reserves without drilling new wells.
MORE VALUE BY REINTERPRETING EXISTING DATA
Conventional characterization depends on new drilling, coring, and laboratory analysis—high-cost processes with limited availability. Our technology utilizes archive data, processing it with physics-guided Artificial Intelligence to estimate petrophysical properties, complementing current methods and reducing the cost of acquiring reservoir knowledge.
| CONVENTIONAL METHOD | PHYSICS INFORMED ML |
|---|---|
| Drilling + coring | Archive data (thin sections, cores, logs) |
| Traditional petrophysical labs | Diagenesis-based HFU estimation |
| Statistical models with hallucination risk | Physically constrained models (hallucination-free) |
| Weeks or months | Days or weeks |
FULL TRANSPARENCY: WE ARE NOT A BLACK BOX
In the oil industry, every decision regarding reserves must be auditable. That is why we incorporate TreeSHAP, a game-theory-based explainability tool that decomposes each prediction into the exact contribution of each input variable.
What does this mean for the operator?
- Traceability: you can see how each data point (thin section, Gamma Ray log, core porosity) influenced the estimated HFU.
- Geoscientific validation: a petrophysicist can confirm that the learned relationships are geologically consistent, free of spurious correlations.
- Regulatory confidence: when submitting a reserve re-evaluation report, you can demonstrate the exact line of reasoning of the model, eliminating black-box opacity.
TreeSHAP allows us to guarantee that the model is not taking statistical shortcuts, but learning the true physical relationships governing the reservoir. The result: accurate, explainable, and auditable predictions.
FROM HISTORICAL DATA TO AN UPDATED MODEL
The solution integrates thin-section data points, well logs, and coring information to train Physics Informed Machine Learning models that estimate petrophysical properties and support the update of the static reservoir model.
Hydraulic Flow Units (HFU): instead of training with conventional petrophysical labels, our model learns to identify HFUs that reflect the real diagenesis of the reservoir, ensuring geological consistency and physical predictability.
BENEFITS FOR THE ASSET
- Uncertainty reduction in property distribution.
- Fast static model update without drilling new wells.
- Development and reserve decisions based on physically validated data.
- Elimination of hallucinations typical of unconstrained statistical models.
- Full explainability with TreeSHAP for audits and geoscientific validation.
INTELLIGENT DRILLING: HYBRID PHYSICS + ML MODELS
Purely statistical machine learning does not work in drilling. What works are hybrid models that incorporate physical phenomena and engineering calculations into the solution.
Drilling data is scarce, noisy, and non-stationary. A purely correlational model learns the field, not the phenomenon. Physical constraints act as a regularizer: the model generalizes with less data and does not produce physically impossible outputs.
US patents granted support this methodology: from geology-guided ROP control to automatic slip detection via image processing.
The Premise
Every well decision must be auditable; a model that respects physics is explainable in engineer language. Physical constraints allow the model to generalize with less data and prevent physically impossible outputs.
Short-Term Applications
- ROP optimization with hybrid models — anchored in the formation's geological response.
- Analytics on historical data — non-productive time (NPT) identification and economic prioritization without new wells.
- Automated event detection — slip positioning, cuttings classification, real-time mud flow measurement.
This technology extends to the drilling domain the same Physics Informed principle that Geo Petro Litix already applies in reservoirs.
Real-time monitoring with edge computing.
Integrated operations center for digital drilling.
CORPORATE TRAINING PROGRAM
Machine Learning Applied to Drilling — The Hybrid Models Approach. A program co-taught by William Contreras (drilling) and Pablo Guillén (machine learning).
The goal is not to teach model coding. It is to teach technical and management teams to distinguish which machine learning applications in drilling are real, which are hype, and what the organization must have to capture value.
5 Modules
- Data baseline — is the operator ready?
- ML Fundamentals & why purely statistical approaches fail — with live demonstrations.
- Short-term application cases — hybrid ROP, historical analytics.
- From analytics to decisions — integration into the real-time operations center.
- The future — roadmap towards closed-loop and autonomous drilling.
What makes it unique?
- The combination of a drilling engineer and a data scientist who runs live models.
- No programming required: it trains judgment for interpreting results.
- Real-world cases with granted patents, presented along with their limitations.
- Deliverable: data maturity diagnostic matrix for the operator.
Duration: 3 to 5 days (in-person or modular). Inquire about corporate rates.
📧 Request program informationAbout Us
Rafael Antonio Díaz
CEO – Executive Leader
PMP. B.S. Petroleum Engineer since 1995
Pablo Guillén
CSO – Scientific Leader
PhD. Machine Learning / AI applied to Oil & Gas since 1995
William B. Contreras
Sr. Digital Drilling Specialist
MSc. B.S. Mechanical Engineering since 1990