Abstract This paper presents a new reservoir modeling technique to improve reservoir characterization. We integrate concepts of multivariate statistical analysis, neural network, and fuzzy logic to get better petrophysical models. The first step in our methodology is to preprocess data, for which we used Q-Q plots with 95% confidence ellipses for quality control. After that, we identified the dominant and the optimum number of independent variables from core and well log using principal components, factor analysis and fuzzy logic concepts. Then, we applied neural networks to model any target variables.1 To check model validity, we used residual plots, average absolute errors, and correlation coefficients. Finally, we used the hydraulic interwell connectivity index (HICI)2,3 to characterize and estimate reservoir connectivity. To calculate HICI, we integrate the neural network petrophisycal models, geostatistic concepts, production performance and reservoir engineering. In this research, we applied the proposed technique to model permeability for the highly heterogeneous cretaceous Caballos formation in Toldado field, Colombia. First, we evaluated the permeability model developed by D. K. Davies4 in this field. His method is based fundamentally on the identification of rock types (intervals of rock with unique pore geometry). We think this a good approach to improve the prediction of permeability. However, Davies gets traditional models (logarithmic of core permeability vs. core porosity) for each rock type. A permeability value from core analysis represents about one cubic inch of rock and may not reflect the permeability distribution in a reservoir. Also, many authors5-8 have found that permeability may be a function more than porosity. The permeability derived from well logging data represents an average that samples several cubic feet of rock. We found a 78% of average absolute error and a serial correlation in a residual plot in the Davies's permeability model for Toldado field. This means that the permeability in this case is not merely a function of the porosity. When we incorporated Davies's methodology and intelligent systems, we significantly improved the permeability prediction from core and well-log data. Applying the concepts of principal components and factor analysis, we found that the dominant independent variables to predict the permeability for Toldado field were porosity, index of the effective photo-electric absorption cross section of the formation (PEF), and gamma ray logs. The permeability prediction from the intelligent system appears to be the best model. The average absolute error is less than 8%. Also, there is no apparent serial correlation, and the mean is around zero in the residual plot for the neural network model. We validated the neural network permeability model by applying the concept of hydraulically connected zones. We used porosity, permeability, thickness and accumulated production rates to estimate HICI. We found that this concept worked well only with permeability-confident models, and it could be used to determine infill wells, damaged or fractured zones, or water zones to be isolated. These models helped us to improve the history match and make recommendations to increase the oil production by about 500 bopd.