Getting More Value of Your Reservoir-Rock-Typing Model based on Artificial Intelligence

Nowadays, petrophysics are using machine-learning models and artificial intelligence that provide the potential to establish multidimensional, nonlinear, and complex models. i.e reservoir rock typing models (RRT. However a common question arise first time when they start to use this methodology. How could be established if a model improve your reservoir description without a dynamic-reservoir-model validation? There is not a unique or straightforward solution. A workflow will be showed but it should be adjusted according to the data available and reservoir under study. If results are not satisfactory at any point it will be more logical to come back and tuning up the model instead of follow verifying next steps.

If each step has been approved satisfactorily confidence shall be higher and any simulation problem could be explained in terms of upscaling process, geostatistical distribution and scarce pressure-production data.

Workflow will be as follow:

1)   Determine statistical parameters of your RRT model based on Artificial intelligence.

2)   Use univariate and multivariate statistics by rock types (histograms, box diagrams and crossplots) at core space, in order to check congruency.

3)   Propagate your model across to the cored well(s) and compare your results with all geological data available (Core photos, XRD, SEM, thin sections, facies, cuttings, and image logs).

4)   Test permeability prediction based on RRT model. A good match between predicted permeability and core data is encouraging.Its a good idea to compare also with permeability wireline-formation-tester derived

5)   Verify your model on blind wells in term of RRT and permeability prediction. Many authors have proposed this step as the primary verification step. In this publication have been decided to delay until the fifth step because incongruences at this point could be explained by the data nature itself instead of a weak model. Any inconsistency at this point should explained and shall be followed with extreme caution.

6)   Propagate your RRT model to uncored wells. Use univariate and multivariate statistics by rock types (histograms, box diagrams and crossplots) at log dimension, in order to detect anomalies. If there is any anomaly in well(s) check log data and corroborate if the well(s) belongs to the reservoir and the same depositional environment.

7)   Scal data must be analyzed by each rock type. Another workflow should be followed in order to deal with wettability issues.

This workflow is rock-type method independent. You can always use it as a guideline when rock types are based on A.I or machine learning. Does not matter if FZI, MICP (Winland or Pittman method), electrofacies, Wooddy-Wright-Johnson or J-Leverett approach, or Lucia Classification is used. You can follow this philosophy in order to decide if you classification is valid.

Please take into account “A.I in rock typing is not only about statistics". All geological data available must be compared, permeability prediction has to be improved, and model propagation on uncored wells should be validated

Finally, a robust A.I method should be a “Two Step Algorithm", at a first step the algorithm must automatically determine reservoir rock types without any “a priori knowledge” and after must be allowed to the user merge any rock type according to the geological knowledge "