Gains:
- Ability to explain the workflow in predicting porosity, permeability and reservoir properties with AI
- Ability to position machine learning in well log interpretation, facies classification and production forecasting
- Ability to verify reservoir outputs through physical boundaries, well testing and uncertainty analysis
An oil or gas reservoir is an invisible treasure trove of fluids trapped in porous rock, miles below the surface. We cannot see it directly; We only "read" it with the logs, cores, seismic shadows and production data we receive from the wells. Reservoir characterization is the task of reconstructing from these indirect measurements how much fluid the rock holds (porosity), how easily it releases fluid (permeability), and what kind of three-dimensional structure the reservoir is. AI is powerful in this area: interpreting well logs, classifying facies (assemblages of rock types), predicting porosity-permeability relationships, and modeling production trends. But reservoirs, like other geological fields, are full of uncertainty, and not every prediction can be a basis for decision without being linked to physical boundaries, well testing and uncertainty analysis.
In this unit we will cover three AI aspects of reservoir study: well log interpretation and facies classification (extraction of rock type and properties from logs), porosity-permeability prediction (prediction of reservoir quality), and production prediction (how much the well will produce in the future). Geostatistics is the backbone here too; We fill the space between wells with variogram and simulation. Note: the oil and gas sector is very commercially sensitive; All samples are studied with anonymous and representative values.
Well Log Interpretation and Facies Classification
Well logs are curves that continuously measure the physical properties of the rock along the bore: gamma ray (clay content), resistivity (fluid type — water/hydrocarbon), density and neutron (porosity), sonic (rock hardness). A petroleum geologist reads these curves together and extracts the rock type and fluid at each depth. This is pattern recognition, and AI is very good at it: consistently sorting tens of thousands of meters of logs into facies.
But there are two limits. First, the log interpretation is a model, not a measurement; The same log may be open to different facies interpretations. Secondly, if the model blindly transfers the log-facies relationship it has learned in one field to another field (different depositional environment, different diagenesis), it makes serious mistakes. Therefore, the facies output is calibrated with core data; The core is the anchor that connects the log interpretation to the ground truth.
Tip: Before trusting the output of a facies model, ask “which wells have been cored and how accurate is the model in those wells?” ask. An uncorroborated facies estimate is an untested hypothesis, no matter how plausible it may seem.
Porosity and Permeability: The Most Critical Distinction
The two keys to reservoir quality are porosity and permeability, and the difference between them is one of the most important points of using AI. Porosity is the void ratio of the rock and can be derived from logs with reasonable accuracy. Permeability is how easily fluid flows through the rock and is not measured directly from logs; but it is estimated indirectly. Permeability is exponentially sensitive to small changes in porosity and depends on pore geometry, cracks, and cementation.
That's why permeability prediction with machine learning is both attractive and dangerous. The model learns porosity in core wells and permeability from other logs; but the relationship it learns is specific to that facies and that pressure-temperature condition. Extrapolation to a different lithology or to an uncored well (moving outside the learned range) will result in large errors. Permeability estimation must be linked to core measurement and well testing (pressure drop/swell tests give actual flow behavior).
Caution: Even if the permeability estimate produces a "good looking" graph, it is unreliable if applied outside the facies and pressure conditions of the training data. The model learns a correlation, not a physical fact; always test with core and well testing.
Production Forecast and Uncertainty
How much a well produces in the future depends on how much fluid the reservoir holds and delivers. Classical methods (decline curve analysis) extend the past production trend into the future; AI enriches this trend with multi-well data and learns from neighboring well patterns. But production forecasting is inherently uncertain: the reservoir is heterogeneous, pressure changes, operating conditions change. Therefore, a probabilistic range (low/medium/high scenario) is presented, not a single "exact" production curve.
Geostatistics also carries uncertainty when filling the volume between wells. Methods such as sequential Gaussian simulation produce many equally likely models rather than a single "average" model; The difference between these models is the uncertainty itself. Artificial intelligence accelerates and interprets these simulations; but reducing the result to a single number creates false confidence by hiding uncertainty.
Three Mini Cases: By the Numbers
Case 1 — Extrapolation error. The permeability model trained on three cored wells at one site gave a prediction that appeared perfect on the fourth well. But this well was in a clayier facies and the model had never seen this interval; well testing showed actual permeability to be about one-fifth of the estimate. The model was seriously mistaken for a “good looking” graph; The well test anchor caught the error.
Case 2 — Facies calibration. In one gas field, artificial intelligence quickly sorted the logs of 14 wells into facies. When compared to three cored wells, the zone that the model classified as a “clean sandstone” was actually a fractured, low-quality unit. With core calibration, the model was retrained and the reservoir volume estimate was brought down to a realistic level.
Case 3 — Script honesty. In a production forecast, the model gave a single curve and expressed ten-year production in odd numbers. When the team generated a low/medium/high scenario through a sequential simulation, they found that the range was very wide (the high scenario was approximately twice as low). This width indicated that additional well testing was required in the investment decision; The single number dangerously concealed this uncertainty.
Weak Prompt / Strong Prompt
Weak prompt:
Look at these well logs, calculate the permeability, and tell me how much the well will produce. [log data]
Powerful prompt:
Your role: Assistant reservoir geologist. Giving SINGLE EXACT NUMBER for permeability and production with the logs below. Instead:1) State which wells can be confirmed by core when making facies interpretation.2) Explain that the permeability estimate cannot be measured directly from logs, under what conditions (facies/pressure) it will be valid, and the need for core/well testing.3) Establish a low/medium/high scenario logic for production, not a single curve.4) List 4 sources of uncertainty that affect the result.Data is anonymous; No real field/company.
Four Copiable Templates
1) Log-facies calibration check:
Explain which wells are cored for facies classification and how we can measure the model's accuracy in these wells. Mark unverified zones as "untested"; Don't trust blindly.
2) Permeability validation:
For permeability prediction: specify the facies and pressure range for which the model is trained; Mark which wells/zones are outside this range (extrapolation). Emphasize that the prediction should not be trusted without core and well testing in these zones. Don't give an exact number.
3) Production scenario draft:
Establishes low, medium, high production scenario logic from decline curve/production data. Explain what assumptions each scenario is based on and what the differences between them mean. Offering no single "exact" curve.
4) Uncertainty simulation interpretation:
Interpret the output of a multiple equally likely geostatistical model (e.g., sequential simulation): how wide is the distribution of the volume estimate, which region is most uncertain, and with what additional data (well, seismic) is this uncertainty narrowed? Reducing to a single number.
Reservoir Specification, Measurability and Verification
feature/output
Can it be measured from logs?
AI risk
verification anchor
porosity
Derived with reasonable accuracy
Calibration drift
Core porosity
permeability
No, indirect guess
Extrapolation error
Core + well test
facies
Comment (model)
Off-site transportation
core definition
Fluid type
I deduce from resistivity
wrong comment
Pressure/sample testing
Production forecast
trend extension
false certainty
Scenario + well test
Hint: For every estimate on the reservoir "what direct measurement could disprove this?" Test it by saying. Core for porosity, well test for permeability, core definition for facies. A prediction that cannot be refuted by direct measurement has not yet been confirmed.
Common mistakes
- Seeing permeability as porosity. Permeability is not measured directly from logs; Extrapolation gives large errors, core/well testing is essential.
- Using the facies model without core. A facies that has not been calibrated by core is an untested hypothesis.
- Carrying the learned relationship to another field. Different depositional environment and diagenesis change the log-trait relationship.
- Presenting production with a single curve. If uncertainty is not given with scenarios, the investment decision is based on false certainty.
- Reducing the simulation to the mean. Multiple models indicate uncertainty; Reducing it to a single number destroys information.
In summary
- The reservoir is invisible; It is read indirectly with logs, core, seismic and production data and every prediction carries uncertainty.
- Porosity can be derived from logs; Permeability is not measured directly, it must be coupled with core and well testing.
- Facies classification is a model; unreliable without being calibrated by core and transported off-site.
- The relationship learned is specific to educational conditions; Extrapolation is the most common and dangerous mistake.
- Production and volume are presented not with a single number, but with a probabilistic scenario and simulation, with uncertainty clearly given.
Application task
Get an anonymized set of well logs (or representative values). Have the model find out which zones are at risk of extrapolation with the "permeability validation" template; Test whether it stimulates a zone by deliberately placing it outside the training range. Then, use the "production scenario draft" template to produce a low/medium/high scenario instead of a single curve and comment on the width of the range in a sentence.
checklist
- [ ] I know the difference in measurability between porosity and permeability and that permeability requires core/well testing.
- [ ] I do not consider facies classification reliable without calibrating it with core.
- [ ] I recognize the risk of extrapolating the learned relationship to another field/condition.
- [ ] I present production and volume with probabilistic scenarios and simulations instead of single numbers.
- [ ] "What direct measurement could disprove each reservoir estimate?" I test it with the question.