GDI (Geological Driven Integration) is a general-purpose, highly flexible, quantitative interpretation software system. Main applications are seismic lateral and volume-based prediction, seismic pattern analysis, AvO module, and rockphysical / petrophysical analysis. The system consists of a number of modules to integrate, manipulate, process and analyse the data. The modules are organised according to functionality in the menus: dGB-GDI, Wells, Seismic, Analysis, and Utilities. (see list below).

- dGB-GDI
System related functionality; survey definition, software settings.

- Wells
All data and tools related to wells: the geological integration framework, the wells, well manipulation tools, and the pseudo well simulator.

- Seismic
Handling seismic data; generating synthetic seismic and log traces, AvO module, and processing, manipulating, and viewing seismic data.

- Analysis
Analysing data; grids, feature sets, sequences, neural nets, segmentation grids, and model probability module.

- Utilities
General purpose utilities; manipulating graphs, running batch jobs, Gassmann equation, display survey information, on-line help.

Integration framework
In GDI, data and knowledge are combined via the Integration Framework (Figure right). The Integration Framework is a user-defined generic description of the subsurface in terms of geological objects. The units are ordered in a tree corresponding to a hierarchical ordering system, and describes the stratigraphical and lithological structure. The framework units are the building blocks for describing different geological models. (Log-) Properties are attached to the individual framework units. The integrated dataset allows us to study inter-relationships at different scale levels defined by the hierarchy of the integration framework. The optimal hierarchy, in terms of stratigraphic detail and attached properties, depends on the available geological knowledge and data, and the study objective.

Unsupervised segmentation
An interesting interpretation technique is the unsupervised segmentation (clustering) of the seismic response around an interpreted horizon. In unsupervised (or competitive learning) the aim is to find structure in the data. The method produces two output grids: a segmentation grid and a 'match' (i.e. confidence) grid (figures below). The segmentation grid reveals areas with similar seismic response. The result must be interpreted in terms of stratigraphical / petrophysical variations at the target level. The GDI software allows us to analyse and quantify the seismic segments using real and/or simulated wells. For the analysis an integrated dataset must exist. The same technique can be applied in 3D mode.

Supervised learning
The supervised approach requires the presence of a representative data set comprising seismic signals with corresponding geological/petrophysical information. The neural network is trained by examples from this data base, and monitored using an independent test data set. One can use real seismic and well data, and/or synthetic seismic and pseudo-well data. Networks can be trained to either classify, or quantify the seismic measurements. The output is either a prediction grid or an inverted 3D volume. (figures below)

To apply this technique only the seismic data, the interpretation horizon and the essential well data must be loaded. Essential well data in this context are: the positions at target level and target quantities such as good -, medium -, poor - reservoir, or porosity, net-to-gross, production rate etc.

Pseudo-wells
If only a limited number of wells is available, or when the existing wells are not representative for the entire area, it is advised to extend the data base by generating pseudo-wells. These stochastic simulated wells are likely representations of the survey geology and are generated based on existing well data, and on geological and petrophysical knowledge of the area. Stratigraphic sequences, as observed in real wells, can be analysed, using e.g. generalised Markov Chains. Wells with different sedimentary stacking patterns are easily detected by this method. The sequence analysis can be used during simulation to generate pseudo-wells with similar stacking patterns.

Pseudo-wells are primarily used in seismic applications. In general, they are used to quantify seismic measurements in terms of geological and/or petrophysical probabilities. For example, in lateral prediction studies pseudo-wells are used to find relationships between seismic features and underlying well features. In seismic pattern analysis they are used to relate the patterns to geological / petrophysical phenomena.

Other applications are sensitivity analyses and feasibility studies. The objective in sensitivity analysis is to get a feel for the seismic response. Well information is varied in a controlled way and the corresponding seismic signals are visually inspected. This is non-quantitative. Feasibility studies, in contrast, are quantitative exercises. We want to establish how far the seismic data can be pushed, i.e. what information can be extracted from the seismic signals and what information is beyond the resolution.
To generate pseudo-wells the GDI-SIMULATE module is required.

AVO
Within the AVO module prestack gathers, stacked responses and AVO attributes are generated. The module is used for:

- unsupervised AVO inversion: matching AVO models to the seismic character using a UVQ neural network;

- supervised AVO inversion: inversion of offset stack data to rock properties using a MLP or RBF neural network trained on modelled data; and

- AVO scenario modelling: modelling the AVO response of simulated wells (generated with the pseudo-well simulator (see image below).

The AVO functionalities are part of the GDI-PRESTACK module.

Feature sets
An important concept in the software is that of 'Features' and 'Feature Sets'. In general, Features are derived quantities from seismic traces (amplitudes, attributes), and/or from well data (average porosity, fluid fill, net-to-gross etc.). In the software Features are stored in Feature Sets. These can be merged, analysed and manipulated in various ways.

For example, we can extract and combine on a sample-by-sample basis numerous features from different 3D seismic cubes (reflectivity, near offset, far offset, acoustic impedance etc). Trained (supervised or unsupervised) neural networks can be applied to such a multi-value Feature Set to yield 3D prediction / classification or segmentation results (e.g. porosity cube, lithology classification cube, or in case of unsupervised networks a seismic segmentation cube). For this purpose, we can convert the porosity log, or any other log, to time by using the sonic log and resampling to the seismic sampling rate using an anti-alias filter.

Stochastic inversion
Another advanced inversion technique, supported by GDI, is a stochastic inversion via the Model Probability Module (MPM). In MPM, simulated pseudo-wells (or real wells as the case may be) are scored at real seismic locations for:

- seismic response and

- geostatistical probability

The combination of seismic - and geostatistical scores yields a 'probability' for each (pseudo-)well at the analysed location. Scores are subsequently analysed to yield expectation and standard deviation grids per quantity.