The objective of this paper is to apply probabilistic neural networks (PNN) to invert the seismic data volume to impedance by training and validating the acoustic impedance logs at the wells.

This problem is addressed by using the technique of cross-validation, in which we remove wells from the training stage and then „blindly?, predict these wells in the validation stage (Herrera et al., 2006). That is, the fit at the wells is excellent, but the underlying model is too complex and does not lead to physically meaningful results away from the well. However, there is the danger that the neural network can become “over trained”. The low-frequency component missing in the seismic is added from sonic logs to assure a more realistic result (Lindseth, 1979). These relationships are applied to seismic data to predict interwell reservoir properties. The inversion requires the initial value of Z to be known but the limited seismic bandwidth constrains this technique. ANNs have the ability of recognize complex, non-linear relationships between seismic attributes and petrophysical data. Artificial Neural Networks (ANN) was introduced to the geosciences community in 1980?s.

The calibration of well logs (high) resolution with 3D seismic data (low resolution) is a challenge while building the comprehensive geological models. The characterization of these reservoirs is a challenge to the geoscientists in terms of resolution and distribution in an area. These are not only thin but wide spread in the basin. These reservoirs have low permeability often act as both the primary source rock as well as the reservoir rocks. The contribution of production from the unconventional reservoirs like shale and coal is increasing everyday to the energy needs of the World. The inversion is performed with the CGG Veritas Hampson-Russell software. PNN has not only helped to improve the vertical resolution but also the lateral variation in the heterogeneity of the reservoir. Post-stack inversion transforms a single seismic data volume into acoustic. Probabilistic neural networks (PNN) are used to invert the seismic data of coalbed methane (CBM) field from northeast Australia to better understand the reservoir properties of the coals sandwiched between sands and shales. Artificial neural networks are the best method to relate the non-linear fit between the borehole parameters to seismic volume parameters to better understand the heterogeneity of the reservoir properties. Geostatistical methods help resolve this problem, but these methods rely only on a linear fit between seismic attributes and reservoir parameters. This problem is due to low resolution of seismic data as well as the lack of proper models that link the seismic data to the borehole data. The seismic guided estimation of reservoir properties away from the wells is a common problem that geophysicists, geologists and reservoir engineers face every day.

Colored inversion hampson russell