Reliable and comprehensive well log data is essential for a reasonable evaluation of the underground rock formations and the in-situ hydrocarbons. This log data has to he interpreted with good statistical confidence to predict the performance of a well. In order to automate the process of log interpretation and hence potentially reduce its cost, a Neural Network (NN) can be trained for interpreting logs with reasonable accuracy. This paper discusses the possibility of applying the Simulated Neural Network (SNN) technology to the art and science of well log interpretation.
The algorithm used to generate the SNN used the Kohonen unsupervised learning technique. Basic principles of the SNN thus created have been considered and the results obtained from the training and SNN log interpretation have been discussed.
Deep offshore wells are currently being drilled using the state-of-the-art MWD and LWD techniques. Both techniques rely on data being transferred rapidly and accurately to the surface and its immediate interpretation. The information acquired from fast data interpretation can be applied to make sound decisions concerning the drilling operation.
In the case of conventional logging, i.e. after the well has been drilled, an expert needs to look at the logs, collect the available information from each of the log curves, correlate and collate all the information and compare it with other data available from the oil field to predict the nature of the rocks and the quantity of the in-situ hydrocarbons. Therefore, both MWD/LWD and conventional logging, can benefit from automating the process of data interpretation and decision making.
Mathematically, the log analyst's thought process of estimating lithologies and hydrocarbon saturations is analogous to determining a global solution over a range of equations represented by the log suite. If each log curve represents an equation, usually we have a system that is overspecified implying non-uniqueness of the solution. Conventional approaches lead to local solutions which may or may not represent the lithology or the hydrocarbon saturations adequately. A global solution usually eliminates alternative local solutions that are not correct. Optimizing global solutions from data obtained from logs generally provide excellent log interpretations.
It can be inferred from the literature that SNNs are very capable tools for determining the global minima associated with non-linear problems in multi-dimensional space provided that the SNN algorithms are correctly implemented (Baldwin et al. [1989:1], Zurada  and Lippman ).