Mary M. Poulton

Neural networks as an intelligence amplification tool: A review of applications

The sophisticated algorithms we use to process, analyze, and interpret geophysical data automate tasks we used to do by hand, transform data into domains where patterns are more obvious, and allow us to calculate quantities where we used to rely on intuition or back‐of‐envelope estimates. But, the crux of the exploration problem is still interpretation—associating abstract patterns with geologic formations of economic value. Artificial neural networks are able to couple the speed and efficiency of the computer with the pattern recognition and association capabilities of the brain to aid the exploration process. The key concept to understand in the application of neural network technology is that they should not be used as an artificial intelligence tool to replace a human interpreter; rather, neural networks are an intelligence amplification toolkit that allows the interpreter to focus on the important information.More than 102 neural network papers have been published by SEG since 1988, and more than 550...

Location of subsurface targets in geophysical data using neural networks

Neural networks were used to estimate the offset, depth, and conductivity‐area product of a conductive target given an electromagnetic ellipticity image of the target. Five different neural network paradigms and five different representations of the ellipticity image were compared. The networks were trained with synthetic images of the target and tested on field data and more synthetic data. The extrapolation capabilities of the networks were also tested with synthetic data lying outside the spatial limits of the training set. The data representations consisted of the whole image, the subsampled image, the peak and adjacent troughs, the peak, and components from a two‐dimensional (2-D) fast Fourier transform. The paradigms tested were standard back propagation, directed random search, functional link, extended delta bar delta, and the hybrid combination of self‐organizing map and back propagation. For input patterns with less than 100 elements, the directed random search and functional link networks gave ...

Neural network pattern recognition of subsurface EM images

Abstract Neural networks are computer simulations of the brains neural functions; as such they perform well on the same types of problems on which humans perform well, namely pattern recognition. Neural networks have shown the capability to learn human speech, read handwritten signatures and recpgnize human faces. Applied to geophysical data, neural networks offer the ability to estimate model parameters in near realtime. A backpropagation neural network was trained to estimate the spatial location (offset and depth) of a target given an image of the electromegnetic ellipticity. Three components of the magnetic field were measured from which the ellipticity was calculated. Theoretical ellipticity images were used for training the neural network; field data were used to test it. The input data representation was important in obtaining results with 10% error or less from the neural network; generally, smaller input vectors yielded more accurate results. Five different representations were examined: the whole image, the subsampled image, trough-peak-trough, peak amplitude and frequency domain. The frequency-domain representation estimated the target locations in the field data with the least error, 0.4% for the offset and 1.5% for the depth. The network was examined for its ability to generalize, to extrapolate beyond the spatial limits of the training set and to ignore discrepancies between synthetic and field data. The generalization from synthetic training data to synthetic test data had errors near 5% for most offset estimates and near 2% for most depth estimates. We considered extrapolation errors satisfactory (10%) up to 1.5 model spacings beyond the limits of the training set.

Neuro-Drip: estimation of subsurface wetting patterns for drip irrigation using neural networks

Design of efficient drip irrigation systems requires information about the subsurface water distribution of added water during and after infiltration. Further, this information should be readily accessible to design engineers and practitioners. Neuro-Drip combines an artificial neural network (ANN) with a statistical description of the spatio-temporal distribution of the added water from a single drip emitter to provide easily accessible, rapid illustrations of the spatial and temporal subsurface wetting patterns. In this approach, the ANN is an approximator of a flow system. The ANN is trained using close to 1,000 numerical simulations of infiltration. Moment analysis is used to encapsulate the spatial distribution of water content. In practice, the user provides soil hydraulic properties and discharge rate; the ANN is then used to estimate the depth to the center of mass of the added water, and the vertical and radial spreading around the center of mass; finally, this statistical description of the added water is used to visualize the fate of the added water during and after the infiltration event.

Analysis of structural homogeneity of rock masses

Abstract The salient features and capabilities of the statistical tests suggested by Miller (1983) and Mahtab and Yegulalp (1984) to investigate structural homogeneity in rock masses are reviewed. The difficulty in making a decision about statistical homogeneity only through the results of any one of these two tests or through equal-area polar plots, is shown through a case study. New interpretations are suggested for these two tests to use them in assigning relative ranks for the strength of homogeneity of different regions in rock masses. An example is given to illustrate how one can use Millers method with new interpretations along with equal-area plots in making decisions about statistical homogeneity in rock masses. The example also showsan attempt to incorporate results of Mahtab and Yegulalps test with new interpretations to the decision-making process.

Locating Buried Objects for Environmental Site Investigations Using Neural Networks

Neural networks were used to develop an interpretation tool that could rapidly detect objects and estimate their depths using electromagnetic and magnetic data gathered during the Dig‐face characterization experiments at the Idaho National Engineering Laboratory. Targets of interest included metallic barrels, concrete, wooden boxes, long beams and pipes of various metals and PVC, and more. The neural networks analyzed the data on a point‐by‐point basis and from profiles, single‐sensor heights, and multiple‐sensor heights. Data points for twenty‐one of 24 targets were identified as being distinct from background. Twenty‐one of 24 targets were correctly classified as conductive or resistive.Object depths were estimated, on average, to within 13 cm of the true depth. The neural networks took approximately seven seconds to interpret slightly less than 1,000 data points at a time, each with six components, using a 90 MHz Pentium computer.

Estimating one-dimensional models from frequency-domain electromagnetic data using modular neural networks

An artificial neural network interpretation system is being used to interpret data from a frequency-domain electromagnetic (EM) geophysical system in near real time. The interpretation system integrates 45 separate networks in a data visualization shell. The networks produce interpretations at three different transmitter-receiver (Tx-Rx) separations for half-space and layered-Earth interpretations. Modular neural networks (MNNs) were found to be the only paradigm that could successfully perform the layered-Earth interpretations. An MNN with 16 inputs, five local experts, each with seven hidden processing elements, and three outputs was trained on 4795 patterns for 200 epochs. For two-layer models with a resistivity contrast greater than 2:1, resistivity estimates had greater than 96% accuracy for the first-layer resistivity, greater than 98% for the second-layer resistivity, and greater than 96% for the thickness of the first layer. If the contrast is less than 2:1, the resistivity accuracies are unaffected but thickness estimates for layers less than 2 m are unreliable. A Tx-Rx separation of 16 m with maximum depth of penetration of 8 m was assumed for the example cited.

Borehole electrical resistivity modeling using neural networks

A neural network approach has been applied to model downhole resistivity tools, i.e., to generate a synthetic tool response for a given earth resistivity model. The microlaterolog (MLL), shallow dual laterolog (DLLs), and deep dual laterolog (DLLd) tools are modeled using neural networks to demonstrate this approach. Efforts have been made to select various neural network parameters, including the type of neural network, the length of input data for training, the number of hidden nodes, and the number of training samples.A modular neural network (MNN) has been selected because it can facilitate the training and prediction of tool responses in formations with large resistivity variations. The input data for training are taken to be the model formation resistivity values sampled over a depth window. The window length is chosen based on the tool lengths. Three different window lengths are used for experiments: 6.1, 9.1, and 30.5 m. We found the longer window lengths generally have higher modeling accuracy fo...

Recommendations for IP Research

The John S. Sumner Memorial International Workshop on Induced Polarization in Mining and the Environment was held 17–19 October 1994 in Tucson, Arizona. The event, dedicated to the memory of an IP pioneer, attracted 175 people from 18 countries. An objective was to get recommendations for geologic, geochemical, and geophysical research from IP users and practitioners for IP applications to mining and environmental problems. Conventional IP, as well as IP effects in EM and GPR, were considered. These measurements include complex conductivity, complex dielectric permittivity, and complex magnetic permeability in the frequency range 10−3–109 Hz in multidimensional earths.

Preprocessing GPR signatures for cascading neural network classification

A solution to the problem of long training times is found by segmenting classification and parameter estimation problems into several smaller problems. Information cascades from one neural network to another with each level increasing the specificity of the problem. To be used most effectively, the networks require small input pattern vectors. Therefore much pre-processing is done to extract information from the GPR records that is germane to the classification. A novel method of extracting and enhancing the target reflection through the use of logical filters is developed. A cascading network is constructed that classifies the type of target as point or plane and then identifies the composition. Finally, the size and location of the target are estimated. The modular nature of the network allows it to train faster, give more accurate results and be easily modified.