Predicting unconventional well logs from conventional logs using neural networks

The Formation Micro Imager (FMI) log provides valuable information to the engineer by giving a visual indication of the near-wellbore environment. This information may be processed to accurately identify porous zones, map fracture planes, and improve the estimate of the formation porosity. The latter is especially true for certain carbonates, where the FMI improves the estimates of the vuggy porosity. The FMI-predicted porosity is called spot porosity. In fact, in one Permian basin reservoir, reserves estimated by spot porosity and confirmed by production were six times greater than the initial estimate predicted by conventional neutron and density logging tools. However, two important limitations should be kept in mind: the expense of running the tool and the inability to run the FMI tool in cased holes. In this paper, we provide a new way of utilizing the information obtained from conventional logs such as gamma ray, neutron porosity, etc. In addition to the routine information gathered from these logs, we used a neural network to predict the FMI response in wells without actually running the tool. In other words, the neural network determined the complex relationship that exists between the conventional logs and FMI logs. We demonstrate the new methodology on an actual carbonate reservoir consisting of 260 wells. Currently, a suite of nine logs is available for most of these wells. These nine conventional logs represent the inputs of the neural network. The relative importance of each input with respect to its influence on the FMI response is obtained using a fuzzy-logic approach. FMI logs are available from four recently drilled wells. The FMI log constitutes the output from the neural network. Using a robust neural network, training was successfully achieved with two wells. The testing performed on the remaining two remotely located wells provided pseudo-FMI logs quite similar to the actual ones. The wells used for testing were never seen by the neural net. For the studied carbonate reservoir, where conventional tools underestimate porosity, the results promise further field development, because of the apparent increase in the estimated reserves.