The past few years have seen increasing interest in the application of machine learning techniques in the industry, specifically in seismic interpretation. Over a clastic Tertiary clinoform interval in the public F3-Netherland dataset, we benchmarked advanced neural network algorithms against standard probabilistic lithology classifications from seismic data, to understand their benefits and limitations, and to check which approach works best under which circumstances.
For the well data interpretation, GR, sonic, density and synthetic shear wave logs from four wells were used. To ensure that we were able to identify electrofacies, which can be predicted by seismically effective parameters, the input for the facies classification was limited to acoustic impedance (AI), Vp/Vs and density, with the Vshale log as lithology control. We grouped the data into three main facies types: clean sandstones (potentially good reservoir), clean shales (non-reservoir) and a transition facies of shaly sandstones and sandy shales (conditionally named silt). Crossplots of the elastic properties show that sandstones tend to have lower impedances and a lower Vp/Vs ratio than the shaly facies, with a significant overlap of the facies types (see figure 1).
For the simultaneous inversion of 8-, 18- and 28-degree angle stacks, low frequency models for Vp, Vs and density were built as an interpolation of the well log data, to compensate for the lack of low frequencies in the seismic data. The results of the inversion were P-Impedance and S-Impedance cubes, from which the Vp/Vs ratio was calculated, and a density cube. The standard probabilistic lithoseismic classification approach is based on probability density functions created for each of the identified lithotypes. Pairs of elastic attributes (like P-Impedance and Vp/Vs ratio or LambdaRho and MuRho) are applied to define the probability of a data point to belong to one of the predefined facies types. The PDFs are usually built from well log data, but it is possible to manually increase the deviation range for each parameter. This gives us the opportunity to cover a broader extent of geologically possible values than the specific range encountered in the wells. It is one of the significant advantages of this approach, especially if there is only a limited amount of well data available.
For the test, we used the P-Impedance and Vp/Vs ratio cubes to create probabilistic facies cubes for sandstones, silt and shales. The resulting geometries, especially of the reservoir facies, fit the conceptual geological model very well, with cleaner sandstones located at the shelf edge and in basin floor fans (figure 3a).
Democratic Neural Network Association
The machine learning algorithm used in the next step of the workflow is a Democratic Neural Network Association from Emerson. It employs several neural networks running in parallel that simultaneously learn from the same data set using different strategies. The outcome of this workflow is a probabilistic facies model that predicts the most likely facies distribution and associated maximum probability, as well as the probability relative to each facies. The DNNA was trained on P-Impedance, Vp/Vs ratio, and density traces extracted at well locations to predict the three facies types. Only the three attributes were used to ensure a fair comparison with the results of the probabilistic lithoseismic classification. Three of the wells were used for training, with one left as a blind test. The output DNNA classification had significantly higher resolution and the facies distinction was sharper with less overlap than in the lithoseismic classification. Again, the cleaner sandstones are predominantly located at the shelf edge and in the basin floor fans (figure 3b).
In other projects, we used significantly more poststack and even prestack attributes, which eventually resulted in a more detailed but still stable classification result. However, in heterogeneous data sets, a larger number of wells is required to provide a statistically valid basis for training, which is one of the obvious limitations of the approach.
One of the benefits of neural network applications is that the inversion can be run directly to predict rock properties like porosity or shale volume from a set of input attributes. Here as well, it is possible to use many physically meaningful attributes. In the presented case, the training data set consisted of the smoothed porosity and shaliness logs, and several seismic attributes extracted along well trajectories: seismic inversion outputs, rotated angle stacks and some frequency attributes. Our experience shows that the employment of absolute inversion attributes in addition to reflectivity attributes, increases the quality of the prediction because prior information like the compaction trend or lateral facies changes can be taken into account. And whereas the traditional way of, for example, porosity mapping, is via (linear) regressions from impedance or density cubes, the use of multiple attributes enables a reliable prediction of nonlinear dependencies at a higher resolution than the input seismic data (figure 2).
To avoid overtraining the NN, where the prediction can become unstable as distance from the wells increases, parts of the training data need to be excluded from the training process and used as a blind test.
For surveys with limited well data input – when there is not enough data to train a neural network properly – the probabilistic classification of lithotypes from seismic data works best. We have the possibility to extend the expected lithotypes and the rock physics parameters beyond the data encountered in the well. However, rock physics models and inversion results must be properly scaled, which can be time consuming. Only pairs of elastic properties can be used for classification (P-impedanceI- Vp/Vs, Lambda-rho-Mu-rho). This is a particular limitation in settings with a significant overlap of rock properties between the lithotypes. The level of detail of the prediction results is limited by the seismic frequencies.
If there is access to enough good well information to train a neural network, lithotypes or even rock properties like Vshale and porosity can be predicted with a higher detail than seismic resolution, even if there are nonlinear relationships between single attributes and rock properties. There is practically no limitation in the number of input attributes. Best results are achieved if (absolute) inversion results are used in addition to reflectivity data. Scaling is done by the neural network, but an accurate well tie is required. The success of the classification depends on the amount and quality of the training data (well data). In any case, blind tests should be used for quality control. If there is not enough well information, it is better to use lithoseismic classification.