Machine learning assisted seismic interpretation: An integrated workflow for structural/stratigraphic interpretation, combined with reservoir characterisation.

Introduction

Interpretation of reflection seismic data has come a long way, from structural travel time interpretation, via stratigraphic amplitude and waveform driven interpretation, to full geobody extraction and quantitative interpretation of rock properties. Irrespective of the purpose of seismic interpretation (e.g. regional exploration to field development studies), the main steps are mapping the structure of the subsurface layers and characterizing different properties of beds. A great deal of effort has been placed on finding new approaches to carry out seismic interpretation in semi- or fully automatic ways. Machine learning (ML) has contributed to more rapid and accurate data analysis in many disciplines including earth sciences. The latest developments in ML algorithms, high performance computing, and open source libraries, have enabled new and different workflows for petroleum data science including seismic interpretation. New workflows are more data driven and efficient, with improved measures of uncertainty.

Method and Data

Different ML approaches have been applied to individual steps of seismic interpretation. We have recently (Larsen et al., 2018) showed the role of ML in greater petroleum geoscience workflows. This means that given the developments made in different ML applications in petroleum geoscience, it is the right time to deliver integrated workflows for interpretation which benefit from ML in all stages of seismic interpretation. We present one workflow which includes a variety of ML algorithms (including fully convolutional deep neural networks) and covers most stages of seismic interpretation including seismic to well tie, structural, stratigraphic, and quantitative interpretation. We have applied the workflow on a 3D seismic survey from Norwegian Barents Sea with one well within the survey. The seismic data is prestack time migrated with 12.5x12.5m bin size. Full stack volume, three angle stacks, and migration velocity field and complete suite of wireline logs were used to carry out the study.  For the well tie, we used ML to correlate synthetic seismic traces with real seismic data. ML-based well tie analysis has not only enhanced the quality of the match but also has reduced the time need to perform the well tie. Using the tied well, we selected three main geological markers. These markers have been labelled (interpreted) on one inline and used to train a deep neural network. Then the model automatically populated the picks over the whole 3D survey. The quality of the automatic interpretation can be further improved by adding more picks or labels. Fault detection and mapping is the next step in interpretation which is time consuming. We applied optimum surface voting (Wu and Fomel, 2018) together with deep convolutional neural network architecture for pixel-wise segmentation (e.g. Badrinarayanan et al., 2015) to detect, and enhance faults in the data. The training sets are few fault picks which provided on selected lines. We have also used similar deep neural network (convolutional encoder-decoder architecture for semantic pixel-wise labeling) to detect and highlight sedimentary features such as fluvial channels of Triassic age which are the main reservoir target in the study area. Seismic full and angle stack volumes were used as features in the training with labels representing different cuts of channels in selected seismic lines. This step can be adjusted based on the geological features of interest (for example karst, clinoform packages, different reef types, sand injectites). The final stage of seismic interpretation is quantitative interpretation, in which conventional approaches are treated as a separate step and seldom integrated with previous stages of interpretation. Seismic inversion based methods are the dominant techniques used to derive quantitative rock and fluid properties from seismic data. However, we applied deep convolutional neural network to train models for rock and fluid properties such as porosity, or lithology directly from partial stack seismic data. The labels for training of such neural network architecture is provided from the well within the seismic survey. 

Conclusion

Using a 3D seismic survey and well data from the Norwegian Barents Sea, we demonstrated a fully ML-based seismic interpretation workflow. The approach covers almost all stages of seismic interpretation and enabled us to carry out seismic interpretation faster. The other important advantage of our proposed approach is the possibility to carry out quantitative interpretation integrated with the structural and stratigraphic interpretation.

References

• Badrinarayanan, V., Kendall, A. and Cipolla, R. (2015). Segnet: A deep convolutional encoder-decoder architecture for image segmentation. arXiv preprint, arXiv:1511.00561.
• Larsen, E., S., Purves, D. Economou, and B. Alaei, 2018, Is Machine Learning taking productivity in petroleum geoscience on a Moore’s Law trajectory?:  First Break, 36, 135–141.
• Long, J., E. Shelhamer, and T. Darrell, 2015, Fully convolutional networks for semantic segmentation: In Proceedings of the IEEE conference on computer vision and pattern recognition, 3431-3440.
• Wu, X., and S. Fomel, 2018, Automatic fault interpretation with optimal surface voting: Geophysics, 83, no. 5, O67–O82.