This article proposes a new approach to build a 3D geological model calibrated to well data using an adaptive neural network taking into account pre stack or post stack seismic behavior.
Building or updating a robust geological 3D model with quantification of the uncertainties for oil and gas recovery is a sensitive process which requires time and high-level scientific expertise.
The current article proposes new alternative ways to help the geoscientist to deliver a less biased geological model calibrated at wells within a shorter time frame.
The main workflow aims at propagating sedimentary and fluid content from wells to a dedicated area using an adaptive neural network taking into account pre stack or post stack seismic behaviors. The principles of the method and the first preliminary results will be applied to a case study.
Sedimentary and fluid content of the penetrated reservoir are precise enough at the well location but have to be upscaled to match the seismic data. The upscaling can be achieved either manually and/or using an automatic classification approach.
When reliable seismic data is available, time migrated pre stack data are used to generate different seismic cubes and by-products for reservoir characterization and delimitation. These output cubes can be pre or post stack data.
The result of the propagation is highly dependent on the quality of the well to seismic calibration.
Despite the importance of the main step mentioned above the current article focuses uniquely on the description of the adaptive neuronal network and the results obtained.
The results will be compared in order to analyze the following subjects:
- the comparison of the automatic model with the current deterministic model produced by the geoscientist using the same input data set;
- the comparison of a new geological model using a pre stack time migrated collection of angle gathers only with both models mentioned above.
The experience acquired in seismic data classification has shown us the limits of clustering techniques when reservoir characterization matters. Standard classification algorithms are often influenced by data density more than data characteristics.
This has leaded us to imagine a classification method that would get the class information at the well location and propagate it at the seismic scale. The objective of our supervised classification method is to show up rock properties through seismic attribute values and well log properties.
The key note of our method is to base the neural network learning on well defined classes and on expert input such as fuzzy training data (Figure 1).
Neural network properties
The relationship established between facies groups and seismic attributes is generally not linear. This assumption leads us to prefer the utilization of a multi-layer neural network (Hastie et al., 2002). The number of neurons by class in each layer is determined by the proportion of samples by class in the training data. We deduce a training set from well facies groups and seismic attributes extracted along well trajectories. As this training data stems from wells, we consider it as robust information.