235 7. Seismic inversion and characterization applied to geothermal energy These seismic attributes dedicated to discontinuity analysis do not only highlight fractures, but they may also be impacted by other effects depending on the algorithm. For example, correlation-based attributes are highly biased by dips or lithology changes, while dip-based attributes are sensitive to both local and broad-scale dip changes. All the attributes quoted here unveil fractures: the fracture image is then redundant, while the perturbations are inherent to each attribute algorithm. While analyzing the relevance of an attribute, and before any combination, it can be post-processed using smoothing or threshold, achieving the best compromise between its accuracy and its noise level to capture the fracture intensity. As a control, fracture patterns can be checked in the seismic, although some of them can be too subtle or too discontinuous to be clearly identified and followed from one seismic section to another. Their orientation and continuity can be validated against the conceptual structural model. Attribute combination can be performed using two main methods (Kumar et al., 2017): • Meta-attributes: it consists in a linear combination of the attributes, weighted by their quality. This is an interpretative method. • Clustering or seismic facies analysis: it consists of an unsupervised machine learning technique, as discussed in the next section. The typical responses such as “faults” or “fractures” are highlighted in the map. Such results will be illustrated in the case study section. 7.3.3 Characterization empowered by machine learning Machine Learning is a powerful tool: • To infer classifications or trends, either using wells only or even mixing well and attribute data using supervised approaches. • To analyze typical responses in seismic data or inversion results, through unsupervised approaches. Supervised approaches The supervised approaches (Discriminant Analysis, Neural Networks, KNN, …), consist in building a predictive model to assess a petrophysical property using seismic attributes, commonly, after inversion, P-impedance, S-impedance, and/or their combination (Al-Emadi et al., 2010). This is the descriptive phase. A second phase, predictive phase, consists in using this model to predict the lithology or facies. These two steps are illustrated by Figure 7.10. In practice, to assess the validity of the model, the prediction is performed on the training sample themselves, before any propagation in 3D. Statistics of good assignments, called restitution, are often used. While considering continuous variable prediction, (multi-variable) regression is used, using the same two-step approach. In this case, the RMS error (RMSE) is preferred to assess the uncertainty associated with the prediction (De Freslon et al., 2020).
RkJQdWJsaXNoZXIy NjA3NzQ=