Design optimum frac jobs using virtual intelligence techniques

Designing optimal frac jobs is a complex and time-consuming process. It usu ally involves the use of a two- or three-dimensional computer model. For th e computer models to perform as intended, a wealth of input data is require d. The input data includes wellbore configuration and reservoir characteris tics such as porosity, permeability, stress and thickness profiles of the p ay layers as well as the overburden layers. Among other essential informati on required for the design process is fracturing fluid type and volume, pro ppant type and volume, injection rate, proppant concentration and frac job schedule. Some of the parameters such as fluid and proppant types have discrete possi ble choices. Other parameters such as fluid and proppant volume, on the oth er hand, assume values from within a range of minimum and maximum values. A potential frac design for a particular pay zone is a combination of all of these parameters. Finding the optimum combination is not a trivial process . It usually requires an experienced engineer and a considerable amount of time to tune the parameters in order to achieve desirable outcome. This paper introduces a new methodology that integrates two virtual intelli gence techniques, namely, artificial neural networks and genetic algorithms to automate and simplify the optimum frac job design process. This methodo logy requires little input from the engineer beyond the reservoir character izations and wellbore configuration. The software tool that has been develo ped based on this methodology uses the reservoir characteristics and an opt imization criteria indicated by the engineer, for example a certain propped frac length, and provides the detail of the optimum frac design that will result in the specified criteria. An ensemble of neural networks is trained to mimic the two- or three-dimens ional frac simulator. Once successfully trained, these networks are capable of providing instantaneous results in response to any set of input paramet ers. These networks will be used as the fitness function for a genetic algo rithm routine that will search for the best combination of the design param eters for the frac job. The genetic algorithm will search through the entir e solution space and identify the optimal combination of parameters to be u sed in the design process. Considering the complexity of this task this met hodology converges relatively fast, providing the engineer with several nea r-optimum scenarios for the frac job design. These scenarios, which can be achieved in just a minute or two, can be valuable initial points for the en gineer to start his/her design job and save him/her hours of runs on the si mulator. (C) 2000 Elsevier Science Ltd. All rights reserved.