101 2. Surface geophysical methods The recent development of advanced deep neural networks (DNNs) has opened the door to a new viable approach for directly estimating reservoir properties from seismic data (Formento et al., 2021). Although this kind of neural network requires a large amount of labelled data to be trained, only a limited amount of real well data is required as synthetic data can be used to augment the training set. Recently introduced theory-guided techniques based on rock physics models can help generate a large training set of pseudo-logs, representative of geologic variations, used to feed the DNN for a prediction of petrophysical properties of geological formations from full stack seismic profiles (Formento et al., 2021). The methodology was successfully applied to improve the understanding of the potential for deep geothermal energy in the south of the Paris Basin (Souvannavong et al., 2024). The available seismic data are limited to 600 km of old 2D lines acquired between 1970 and 1990 and 10 old wells which had an available set of Caliper, Gamma Ray (GR), compressional sonic (DTP), density (RHOB), neutron (NPHI) and resistivity logs. The old seismic lines were reprocessed. More than 800 pseudo wells were generated to account for possible geological changes within the 3 reservoir units (Oxfordian, Dogger and Trias), allowing to have a large training set to feed the deep neural network, for a better prediction of total porosity (PHIT) and volume of clays (VCL) from full stack seismic. Figure 2.35a shows a seismic line passing through one of the wells with color-coded reservoir intervals (blue for Oxfordian, purple for Dogger and red for Trias). The estimated PHIT and VCL sections are shown in Figures 2.35b and 2.35c. Figure 2.35d shows the match between the recorded and synthetic seismic traces. On this figure, from left to right: the AI log is displayed in grey to show the acoustic contrast between layers. Then the comparison between the synthetic (black) and recorded seismic trace (red) shows a satisfactory match within the Oxfordian and Dogger intervals but a relatively poor one for the Trias where seismic signal-to-noise ratio (S/N) is lower. Figure 2.35e shows the match between the estimated attributes and the well logs. The predicted attributes (red traces) match well to the log data in general. In the Oxfordian interval, the predicted porosity correctly captures the layer with high porosity at the top of the reservoir (blue arrow on Figure 2.35e). This study illustrates how rock physics-guided deep neural networks were used as a practical alternative to derive accurate total porosity and volume of clay attributes for two carbonate reservoirs (Oxfordian and Dogger) and a clastic reservoir (Trias) directly from full-stack seismic and limited well data (Souvannavong et al., 2024). Reinsch et al. (2017) demonstrated that temperature influences seismic velocities significantly. Du et al. (2024) have studied the Influence of temperature on the velocity-porosity relationship, with laboratory measurements on geothermal core samples. Laboratory measurements have shown that P-wave velocity continually decreases with increasing temperature. This trend in seismic velocity with temperatures is related to microfractures. Using the temperaturedependent Kuster-Toksöz equation (Kuster, 1974), it is suggested that the presence of fluid and microfractures can reduce the effective elastic properties of rocks (Du et al., 2024).
RkJQdWJsaXNoZXIy NjA3NzQ=