statistical multi-attribute neural network inversion | quantitative interpretation
Statistical Multi-Attribute Neural Network Inversion
In some cases, a deterministic impedance based inversion process may not be suitable. Reasons for this could include:
- Poor quality well and/or seismic data. e.g. high random noise content; weak reflectors..
- Lack of resolution e.g. deep targets with insufficient vertical or lateral resolution and aperture.
- No separation of rock and fluid properties in the elastic domains. (Modeling can help predict this prior to beginning an expensive AVO or Impedance inversion.)
In such cases, we offer a multi-attribute neural network inversion workflow, designed to provide seismic facies descriptions calibrated to well-data. Workflow 
elements include:
Attribute Computation (ATTRIB3D)
A variety of seismic attribute volumes will be computed. We generally chose Hilbert attributes of envelope, frequency and phase. Relative acoustic impedance may be added, plus seismic waveform attributes. Pre-computed (client supplied) attributes may also be included.
Classification (LITHANN)
The individual attribute volumes are combined into a single output using an unsupervised neural network classifier. A 2D rectangular Kohonen topology is invariably employed.
Calibration
Calibration of classes is performed at well-locations where known facies are sampled. The output class from the previous phase are combined and re-numbered to yield a calibrated seismic facies volume
Interpretation and Visualization
The resulting 3D seismic facies volume is examined within a suitable 3D visualization system. Geo-bodies may be generated and used for further attribute interpretation.