Purpose: Predict rock properties from seismic information and well data. Theory: A supervised neural network is trained on examples (seismic + logs) extracted along a well track. Typically, input is (absolute) acoustic impedance and/or elastic impedance. Typical outputs are porosity logs, lithology logs (Vshale, gamma ray, lithology codes), and saturation logs. Software: OpendTect + Neural Networks
Workflow:
Open the attribute set window and specify the seismic attributes that you wish to extract along the well track (see Tips below).
Open the Neural Networks window and Select Property Prediction. In the new window specify the input attributes, the target well log, the wells, the well interval, the Location selection (see Tips below), the log type (values or binaries) and the percentage to set aside for testing the network during training.
Crossplot the target values against each of the input attributes. If need be remove / edit points.
Balance the data. This step ensures that each bin in the training set has the same number of examples, which improves training.
Train the neural network. Stop where the test set has reached minimum error (beyond that point overfitting occurs: the network learns to recognize individual examples from the training set but looses general prediction capabillities). Store the trained neural network.
Apply the trained neural network to the seismic data in batch: Processing - Create seismic output, or on-the-fly: right-click on the element in the tree (e.g. part of an inline).
Tips:
To compensate for mis-alignment problems ,you can extract a small time-window from the input (e.g. acoustic impedance volume). To do this use the Reference Shift Attribute and construct an attribute set that extracts at each point e.g. the AI values at -8,-4, 0, 4, 8ms.
Use the "All corners" option for the Location selection parameter to compensate for uncertainties in the well track and to increase the statistics. Each example is then extracted along 4 well tracks that run along the corner grid points of the actual well track.