Application of Artificial Neural Networks for Well Logs

Abstract

Neural Networks are a powerful tool -For computation when the data available is less than adequate.Can solve fundamental problems such as formation permeability prediction from the well log response with high accuracy.Have great potential for computing results from historical data which would otherwise be irrelevant for analysis. Neural Networks find various applications in the Petroleum industry - Optimize the hydraulic fracture design, Permeability Predictions, Facies classification etc. This study deals with the development of a neural network for predicting well log response. This network is trained by input-output pairs of known well log data, using which an estimation model is created. This estimation model can then compute permeability output for input well log data of different offset wells. The neural network created uses a feed forward model with a Levenberg-Marquardt learning algorithm. Error is calculated using Mean Squared Error (MSE) technique. An optimum number of neurons in input, output and hidden layers were set. The network was trained with given pair of input response of well log dataset and its output permeability response was predicted. The study also considers two different cases to mark the importance of optimized training of a neural network before it can be used for predicting permeability values in offset wells. This technique is completely data driven and does not require priori assumptions regarding functional forms for correlating permeability and well logs (Sharma et al. 2011). The network developed in this study gives highly accurate results. Also it is very flexible as the training algorithm, number of layers and other parameters can be changed according to the data set available. It proves to be cost effective and saves time since it does not require core analysis. Neural Networks prove to be a useful emerging alternate tool to conventional methods. Introduction Reservoir characterization is a very important domain of petroleum engineering. It includes description of the spatial distribution of rock properties such a porosity, permeability and saturation, in high detail. This can be achieved from a variety of sources: well logs, cores, seismic, production tests, drill stem tests. Porosity is defined as the fluid storage capacity of porous media. It mainly depends on two factors: grain size distribution and sorting index. Porosity is measured using three techniques namely, Well logs (Sonic, Neutron-Density and NMR), Seismic and Core analysis. Permeability is the measure of the ease with which the rock will permit the passage of fluids. Porosity alone is not sufficient to make a reservoir producible. For the flow of hydrocarbons, the reservoir must have well interconnected pores. Knowledge of permeability is significant for developing an effective reservoir description and quality. Permeability determination is an active research area in petroleum industry as there is no direct formula for calculation of permeability from logs. The conventional methods for permeability calculation are core analysis and drill stem tests. However, these methods have some limitations as they are very expensive and all wells in a field may not have core data (Abdideh 2012). In heterogeneous reservoirs classical methods face problems in determining accurately the relevant petrophysical parameters. Applications of artificial intelligence have recently made this challenge a possible practice (Hamada et al. 2010). Since Permeability and Porosity are typically distributed in a spatially non uniform and non-linear manner (Verma et al. 2012), we propose the use of Artificial Neural Networks for improved porosity and permeability estimation.

Neural model to estimate permeability from well logs and core data

“…Therefore, the training process can be defined as the way a neural network modifies its weights in response to input information. This process is reflected in the modification, destruction, or creation of connections between neurons (Kohli and Arora, 2014;Al Khalifah et al, 2020). The ability to accurately compute functions of unseen inputs by training over a finite set of input-output pairs is referred to as model generalization.…”

“…Geologists, petro-physicists, and engineers require vast sets of variables to activate models or to establish simulations regarding production rates, reserves, and recovery factors (either in completion or site stimulation). Among these variables, permeability (k), is one of the essential characteristics of the rock that considerably influences the decision-making about field development and impacts the management programs (Helle et al, 2001;Singh, 2005;Kohli and Arora, 2014;Eriavbe and Okene, 2019). Permeability is usually measured from core samples in laboratory tests, being this the most reliable method, however, this is not the solution to parameterize bulky environments since extracting too many cores may be prohibitive (due to financial or physical complications).…”

“…Using as inputs gamma ray, bulk density, and induction resistivity, the authors predict k with sufficiency against selected semi-empirical equations. Other researchers have worked in the same direction (Mohaghegh et al 1997;Helle et al, 2001;Habibian and Nabi-bidhendi, 2005;Singh, 2005;Maslennikova, 2013;Kohli and Arora, 2014;Al Khalifah et al, 2020) using practically the same inputs than Mohaghegh et al (1997) with slightly larger training sets. However, due to the nature of the training sets used in these works, the resulting models cannot be applied directly to heterogeneous conditions and their inputs/output ranges are still small-scale.…”

Identifying gaps in using artificial intelligence to support students with intellectual disabilities from education and health perspectives

“…While applying AI techniques to identify dyslexia can often be complex, the preliminary results of recent studies have been satisfactory in leading the way for diagnosis. Kohli and Prasad (2010) used artificial neural networks (ANN) to detect dyslexia, the first systematic approach of its kind. Most recently, Rezvani et al (2019) presented a neurobiological (EEG)-based classifier to diagnose dyslexia for students in grade 3.…”