Gains:
- Ability to explain how the 3D distribution of reservoir properties (porosity, permeability, water saturation) is modeled with AI and geostatistics
- Ability to locate feature spread with facies modelling, variogram, kriging and machine learning
- Ability to verify reservoir model outputs with material balance, well testing and uncertainty scenarios
To fabricate an oil or gas field, we first need to build a 3D model of it: where the reservoir rock is, how porous it is (porosity), how easily fluid flows (permeability/permeability), and how much of the pores are filled with hydrocarbons (1 − water saturation). But we can only measure these features at the points where we drill wells; The huge volume between the wells is empty and must be filled. Reservoir characterization is taking point measurements in wells and extending them to the entire volume using geological logic and statistics. In this unit, we use artificial intelligence and geostatistics (the branch of statistics that models spatial data with uncertainty) in this dissemination work. The unchanging principle: every well-to-well value is an estimate, not an actual one; Uncertainty is part of the model and is tested by physical balance.
Basic Concepts
- Facies: A rock type with unique properties that was formed in a particular depositional environment (e.g. river channel sand, marine shale, reef limestone). Reservoir quality is closely tied to facies: the channel drains sand well, but not shale.
- Variogram: A measure of spatial continuity that describes how feature similarity decreases as the distance between two points increases. It is the basic tool of geostatistics.
- Kriging: A method that estimates the value at unsampled points using the variogram, along with the estimation uncertainty.
- Stochastic simulation: Producing many equally likely models (realizations) compatible with the data, rather than a single "best" guess; It makes uncertainty visible.
- Material balance: Conservation relationship between the fluid produced from the reservoir, pressure drop and remaining volume; physical anchor that independently tests the model volume.
Hint: The reservoir model is an uncertainty distribution, not a single number. The correct language is "P90=35, P50=50, P10=72 million barrels", not "Reserve 50 million barrels". Ask the AI for a scenario/distribution, not a single point estimate.
How AI and Geostatistics Work Together
Classical geostatistics (kriging, variogram) is powerful but struggles with multivariate, nonlinear relationships. This is where machine learning comes into play:
- Facies classification: Assigning facies from the log set (supervised learning); Then distribute the feature using facies as a guide.
- Attribute estimation: Estimating inter-well porosity from seismic attribute + log relationship (seismic driven modelling).
- Gap filling: Completing missing log ranges reasonably from neighboring wells and relationships.
But machine learning doesn't know geological logic. The porosity map produced by a model is incorrect if it is inconsistent with the depositional environment (e.g. high permeability in shale facies). Geological plausibility is the first filter of AI output.
Step by step: Reservoir model workflow
- Structural skeleton. Establish the geometry of the reservoir with horizons and faults from seismic (output from Unit 3).
- Facies model. Determine facies in wells (log + core), distribute to volume with variogram/geological trend. AI speeds up classification.
- Feature deployment. Distribute porosity, permeability, water saturation within each facies by kriging or machine learning; If there is seismic, get a guide.
- Uncertainty. Generate a large number of realizations with stochastic simulation; P90/P50/P10 volumes are output.
- Verification. Test model volume with material balance and history match; Compare with well test permeability.
Three Mini Cases: By the Numbers
Case 1 — Value of facies guide. When porosity was dispersed at a site directly by kriging, high porosity “leaked” into low-quality shale zones. When the facies model was first established and porosity was distributed within the facies, the model became realistic and the stochastic volumes narrowed: P50 volume decreased by 12%, but the uncertainty band narrowed by 40%. The correct structure yielded both more realistic and safer results.
Case 2 — Extrapolation trap. The machine learning model learned the porosity-seismic relationship well in the region where wells are concentrated; but when applied to an edge block without a well, it fell outside the training range (extrapolation) and produced an unrealistic porosity of 28%. The geologist showed that that block was in a different facies zone; the model was invalid in that region. The extrapolation limit is marked.
Case 3 — Material balance conflict. A rich 3D model P50 reserves yielded 68 million barrels; however, the field's early production pressure drop indicated a system with at most ~45 million barrels in material balance. The contradiction came because the model assumed disconnected (isolated) sand bodies to be too connected. The connection was reconsidered and the model was made smaller. The physical anchor fixed the geometric model.
Weak Prompt / Strong Prompt
Weak prompt:
The reservoir model and reserve are derived from this well data.[data]
Powerful prompt:
Generate a reservoir modelingDRAFT plan from the (anonymized) well and seismic summary below. Rules:- First build facies model; Distribute porosity/permeability WITHIN facies. Hydrocarbon feature leaching into shale/low quality facies.- Justify variogram parameters (direction, range) from geological trend, not from data; Write the assumptions CLEARLY.- Present the result as a P90/P50/P10 volume distribution, not a single number.- Flag the risk of extrapolation in wellless regions.- Write how to verify with material balance/production history.Data: [summary]
Four Copiable Templates
1) Facies classification draft:
Suggest a facies classification approach from the following set of logs. Write down the logs to be used, class definitions, and the need for core labeling. Specify the risk of confusion between classes and the geological plausibility check. Data: [log set]
2) Variogram justification:
When establishing a variogram for this porosity data, relate the direction and range selection to the direction of geological precipitation. Explain the risk of assuming isotropic. State the limit of variogram reliability with little data. Data: [porosity]
3) Uncertainty scenario:
Generate P90/P50/P10 volume distribution for the following volume inputs (area, thickness, porosity, saturation, conversion factor). Ask for the uncertainty range of each input, don't assume it. Rank the most influential source of uncertainty (tornado).Input: [values]
4) Material balance cross check:
Compare the following 3D model volume with the given early production and pressure drop data for material balance. If there is incompatibility, list possible reasons (connection, aquifer support, volume error). Model: [volume], production/pressure: [data]
Reservoir Model Verification Layers
layer
What tests
Method
Geological plausibility
Facies-feature consistency
Expert review
well data
Spot measurement compliance
Log/core comparison
well test
Actual permeability/pressure
Pressure transient test
Material balance
volume conservation
Pressure-production relationship
Production history
dynamic behavior
History matching
Common mistakes
- Facies skipping. Distribute features without facies guide and leak hydrocarbons with thing.
- Single point prediction. Reporting a single number of reserves instead of an uncertainty distribution.
- Extrapolation. Blindly applying the model to blocks outside the training data range.
- Skipping material balance. Not testing the 3D model volume with physical balance.
- Leave the variogram to automatic. Choosing direction/range without geological trend.
In summary
- Reservoir characterization is the extension of point measurements in wells to the entire volume with geology and geostatistics.
- AI and machine learning are powerful in facies classification and feature prediction; but geological plausibility is the first filter.
- Distribute features within facies; Report uncertainty as a P90/P50/P10 distribution.
- Mark the extrapolation limits; Do not blindly use the model outside the training range.
- Material balance and production history are physical anchors that test the geometric model.
Application task
Take porosity/facies data from several wells (representative) of a field. Have a modeling plan generated with the powerful prompt. Then: (1) check whether the porosity is distributed within the facies, (2) ask for the reserve as P90/P50/P10, not a single number, (3) roughly provide the volume given by the model with a simple material balance (produced + remaining) logic. Add an extrapolation warning for at least one wellless region.
checklist
- [ ] I understand that the reservoir model is a well-to-well estimate and carries uncertainty.
- [ ] I distribute the features with the facies guide, I do not leak them into low quality facies.
- [ ] I report the reserve as a P90/P50/P10 distribution, not a single number.
- [ ] I mark the extrapolation limits.
- [ ] I verify the model volume with the material balance and production history.