Neural Network Interpretation Of Em39 Well Log Data

We used a Modular Neural Network(MNN) to invert well logging curves from a Geonics EM39<br>induction logger In the interpretation scheme, there are several subsets of networks that depend on the<br>relative resistivities of adjacent layers, e.g. Rl>R2, Rl<R2>R3, etc. The well logging curves are<br>subdivided into several pieces and run through each sub-network. The results are estimates of resistivity<br>and thickness of every layer. Using synthetic data, several training sets were made. The networks were<br>also tested on field data. The networks were examined for their ability to compute the right output patterns<br>for the corresponding input patterns of the training set and the ability to interpret the new patterns that are<br>not present in the training sets.<br>The results show that neural networks do facilitate interpretation of well logging data. When tested on<br>data from shallow wells 4-5m deep representing uniform material, the trained networks had an overall<br>accuracy of about 90% for both resistivity and thickness. When tested on a multi-layer case, the networks<br>gave reasonable estimates for each layer’s thickness and resistivity, although a shift in depth was<br>observed for some layers. Recently, we generated more training patterns for thinner layers and<br>incorporated a new way to pick data points for the input patterns. We found the results more satisfactory.<br>For two complicated multi-layer field cases, the networks had an overall accuracy of 92.1% for the<br>resistivity and thicknesses of the layers.