Date of Graduation


Document Type


Degree Type



Statler College of Engineering and Mineral Resources


Petroleum and Natural Gas Engineering

Committee Chair

Shahab Mohaghegh

Committee Member

Kashy Aminian

Committee Member

Sam Ameri

Committee Member

Mehrdad Zamirian


The primary purpose of this thesis was to confirm the capabilities of artificial intelligence and machine learning through Top Down Modeling in history matching and predicting the oil, gas, and water production rates, reservoir pressure, and water saturation, of one limb of an anticline with water and gas injection. Several other characteristics were also applied to make the model more realistic to industry standards. The second purpose of this thesis was to determine the minimum amount of training and calibration data required in order to obtain good results for this particular dataset by increasing the blind validation in one year increments with each additional model. The aforementioned task was accomplished by first creating the described reservoir in a numerical reservoir simulator (NRS) in order to synthetically generate the dataset for the Top Down Model. Synthetic data from a NRS was desired because all of the details and values were known for the entire reservoir at all time steps. The numerical reservoir model, a unique scenario from any other reservoir built to test Top Down Modeling, had the following list of characteristics to make the model more realistic and unique from models used previously:

 One limb of an anticline with 57 production wells and 20 injection wells brought online in phases

 Water and gas injection

 Random daily and monthly shut-in dates for all production wells

 No communication between some layers

 Presence of sealed faults with no surrounding aquifer

 Water cut operating constraint and partial completion varying with time for all producers

After the NRS model was completed, the data was exported, formatted, and calculated for import into a software to begin the process of creating Top Down Models (TDMs). An iterative process to select attributes, train data-driven models (DDMs), and evaluate the training results was performed for every DDM generated. Variations of attributes selected for training were used with each DDM in an attempt to achieve the best results possible for each blind validation scenario. The length of blind validation and random partitioning of the training and calibration data changed with each model, thus making the iterative process necessary. The TDMs were built from the five DDMs and the accuracy of the predictions made were analyzed on an entire reservoir and individual well basis. If these results were determined to be unsatisfactory, the process was repeated from the selecting training attributes step. As detailed in the results and discussion section, 7 TDMs with varying lengths of blind validation were generated and all displayed acceptable predictions. It is thus shown that AI and ML was successful in history matching and predicting entire reservoir and individual well behavior. The maximum amount of blind validation without predicting new wells was also successfully implemented and satisfactory predictions were obtained. Consequently, one years worth of data for training and calibration was sufficient to teach the TDM recompletion changes and new well behavior to in turn receive adequate predictions.