How to improve your predictions?
Emerge is a powerful geostatistical module from HampsonRussell of CGG GeoSoftware. It can predict log property volumes from well logs and attributes from seismic data.
Besides applying Emerge to predict a volume of log property, Emerge can increase the resolution of the inversion results. Also, it can also be used to predict missing logs by using existing logs that are common to the available wells.
This article has been written by Tanya Colwell, CGG GeoSoftware Product Strategy Manager Reservoir Characterization, and it describes a series of advanced options that you can use in order to improve your Emerge training and obtain more accurate and better final predicted results. To download a PDF copy of HampsonRussell Emerge Best Practices, please click here.
Initial Emerge Considerations
Emerge training should include at least three wells. We recommend adding more well curves if possible. Since Emerge assumes that the target log is noise-free, you must edit the target logs before applying Emerge. For the time domain predictions, since Emerge will be correlating the target logs with seismic data on a sample by sample basis, the proper depth-to-time correlation is critical. For this reason, check-shot corrections and manual correlation are usually necessary.
You can bring in more external attribute volumes. For example, if you have pre-stack seismic, you can bring in the near-stack, mid-stack and far-stack volumes, flagging one of them as “seismic” and the other two as “external attributes”. You can use AVO attributes, inversion output volumes, LMR volumes or any other attributes such as Coherence, Curvature, etc.
For the depth domain volume predictions, note that HRS10.1 Emerge does not use the active well depth to seismic depth mapping table. This functionality will be added in HRS10.2.
Shift Target Logs
Sometimes the correlation can be improved by applying residual time-shifts to the target log relative to the attribute. Run a single attribute analysis. Determine the best single attribute. Apply the residual time/depth domain shifts to your target logs with respect to this attribute. Rerun the single attribute training. The correlations should improve.
Exclude Outlier Points
Remove the outlier points by using exclusion zones on the Emerge cross plot. After the points are removed, they will be marked with gray squares on the target logs and will not be used in the subsequent Emerge training.
Test various convolutional operator lengths under the Multi-Attribute List training. Using the Convolutional Operator is like adding more attributes: it will always improve the Prediction Error, but the Validation Error may not improve – the danger of over-training increases. As the operator length increases, the Training Error always decreases. The Validation Error decreases to a minimum and then increases again for longer operators.
In the example below, the validation plot on the right shows that the optimal operator length was a 7-point operator with three attributes.
Train within the same geological zone
For volume predictions, when you do not need to predict the zone over the entire domain window, we recommend narrowing the training interval. Emerge will give better results if the zone has consistent lithology.
If multiple zones are present, you can use an “indicator” log to exclude certain intervals with different lithology.
Geologically Meaningful Attributes
Emerge is a statistically based program, so sometimes it is difficult to bring in geological meaning to some of the attributes that are selected by Multi-linear regression training.
If you doubt that a specific attribute affects the results, you can exclude it from training.
For example, Time, X-Coordinate and Y-Coordinate are not selected for the training below:
Apply Internal Attributes Option
If you use more than one seismic volume, e.g. near/mid/far stack, then please select an option to apply internal attributes (such as Amplitude Envelope, Derivative, etc.) to the search. However, you do not need to apply these additional internal attributes to the already complicated “external” attributes, e.g. P-Impedance inversion or LMR volumes.
Use Bandpass Filter Option
In the attribute list, we have a series of default frequency bandpass filters, ranging from 5 Hz to 70 Hz. Nowadays, some seismic surveys (i.e. oil sands) may contain frequencies higher than 100 Hz. Therefore, our default set of frequency bandpass filters may not be enough to include all the frequencies of the seismic data. Emerge allows us to define a new set of frequency bandpass filters instead of the default ones.
Filter Target Logs to Seismic Frequency
For volume predictions, you may filter target logs to seismic frequency bandwidth. However, note that filtering the target logs will always produce an improvement in the numerical correlation result, but that may be achieved at the expense of the resolution of the output.
Neural Networks Training
Compared to step-wise regression, Neural Networks can enhance the high frequency resolution.
Start Neural Network training with the best Multi-linear regression list you have identified. Choosing “yes” here means that the neural network will have exactly the same attributes and the same operator length as the selected multi-attribute transform.
Probabilistic Neural Network
Try PNN with a cascading option, since sometimes it can give a better result. With the cascading mode, the first calculation that the network performs is the multi-linear regression with the same four attributes. The predicted log from that calculation is then smoothed with a smoother length given on the Neural Network training dialog. The PNN Neural Network is then used to predict the residual (the high-frequency component of the logs that is not contained within the smooth trend). You then get the final predicted log by adding the trend from the multi-linear regression and the predicted residual from the Neural Network.
Radial Basis Function
Try RBF as a neural network. Because the RBF network is an exact mathematical interpolation scheme, the training data will be optimally fit. For small training datasets, the RBF network may give a higher frequency result than the PNN. Also, the RBF network can run considerably faster than the PNN.
Multi-Layer Feed Forward Neural Network
If you are aiming to fit Emerge predictions to the wells better, we suggest a more aggressive use of Neural Networks. Maybe try MLFN with many nodes in the hidden layer and a significantly large increase in overall iterations. The run-time would go up and the validation error would not necessarily be lowered, but that could be offset by other benefits, like fitting the target data better and having greater geological continuity.
Download a copy of HampsonRussell Emerge Best Practices: Download (PDF, 1MB)