Fluid classification is a critical factor in decision of reservoir and production problems. Reservoir fluid can be classified into five types according to laboratory and production data as black oil, volatile oil, gas condensate, wet gas and dry gas. In this work a novel application of Neural Networks (ANN) is presented. Based on production and laboratory data neural networks model is developed for automatic classification of reservoir FLUID. More than 450 samples of five types of reservoir fluids are used to develop the neural network model. About 70 % of data are accepted for neural network training, 15 % for validation and 15 % are used as test set. The importance of different input fluid properties in classification was studied.
The different types of architectures for different groups of input data were tested to select the optimal neural network architecture by fitness criteria. The optimized neural network model was capable of classifying the reservoir fluids with high accuracy. The performance of ANNs models was determined by classification quality index and network error.
The model has been applied successfully to classification of Yemeni fluids using different range of parameters. The results show that the proposed novel ANN model can achieve high accuracy.

Equilibrium constant has many applications in solving problems in reservoir engineering and petroleum processing. Various correlations are available for estimating K- values for heptanes plus fractions. These correlations can be classified into simple and complicated. However these correlations are not able to predict K values adequately for a wide range of conditions. They lose validity in specific range of pressure and temperature and exhibit some error. In this work neuro-fuzzy modeling techniques (ANFIS) is developed to predict K- values for heavy fractions. A large collection of K- values data points (more than 1340 data points) were extracted from experimental 570 PVT reports using the principal of material balance are used in developing the neuro- fuzzy model. 80% of the data points were used to train ANFIS model and 20% of data sets were used to validate, and test the model. Statistical analysis (average absolute percent error, correlation coefficient, standard deviation, maximum error, minimum error, etc.) is used for comparison the proposed model with empirical correlations. Graphical tools have also been utilized for the sake of comparison the performance of the new model and experimental data. Results showed that the new hybrid neural fuzzy model outperforms some available empirical correlations.