Geostatistical 3D Reservoir Modeling of Mississippian St. Louis Carbonate Reservoir Systems, Kansas

Kansas Geological Survey

Open File 2004-26


Neural Network and Lithofacies Prediction


Determined predictor variables (GR, ILD, ILM, PE, NPhi, DPhi) and lithofacies categories were used to build Neural Network models. The size of network and damping parameters were optimized using cross-validation and repeat testing with whole training data and randomly chosen partial data sets.


Predicted Lithofacies Scorecard (Counts) for 10 key trained wells with PE Curve


The results of model with network size (number of hidden layers) 35 and damping parameter 0.05 for 10 key wells with PE log. The absolute accuracy for this model is 91%. For a few wells without PE curve, the results decreased but an absolute accuracy of approximately 85% is still attained.



Cross validation performed on different network size and damping parameters with objective function and mad parameter, which is mean absolute difference between the predicted and actual facies number. Neural network model (size=35, damping =0.05) was selected for facies prediction.

Last Modified December 2004