Artificial neural networks have been proposed for solving many problems in the oil and gas industry, including seismic pattern recognition, permeability and porosity prediction, identification of lithofacies types, prediction of PVT properties, estimating pressure drop in pipes and wells, and optimization of well production. However, the technique suffers from a number of limitations. This paper introduces Abductive Networks as an alternative modeling tool that avoids many of the neural networks limitations. While the processing elements in neural networks are restricted by the neuron analogy, abductive networks use various types of more powerful polynomial functional elements based on prediction performance. Based on the self-organizing group method of data handling (GMDH), this technique uses well-proven optimization criteria for automatically determining the network size and connectivity, and element types and coefficients for the optimum model, thus reducing the modeling effort and the need for user intervention. The abductive network model automatically selects influential input parameters and the input-output relationship can be expressed in polynomial form. This enhances explanation capabilities and allows comparison of the resulting data-based machine learning models with existing first principles or empirical models.
The main objectives of this paper are to introduce abductive networks and outline their advantages over neural networks. Also, to put abductive networks into perspective from petroleum engineering point of view and encourage petroleum engineers and researchers to consider them as a valuable alternative modeling tool. The technique was successfully used in areas such as weather prediction, energy forecasting, spectrum analysis, and medical informatics, but has not been much utilized in the oil industry. To demonstrate the usefulness of the technique in this area, we describe the use of abductive networks for predicting the Pressure-Volume-Temperature (PVT) properties. Results indicate that abductive network models outperform other models and empirical correlations. Finally, some other potential applications in the oil and gas industry are briefly discussed.