Jan Limbeck (Shell) opined that conventional workflows suffer from poor scalability in the face of big data. Shell has applied machine learning to seismic interpretation in its in-house developed GeoDNN deep neural network. Conventional interpretation workflows are based on ‘siloed,’ semi-automated steps. Physics-based models may not be designed for the business at hand, the process takes too long, and uncertainties may go unrecognized. GeoDNN performs ML-derived feature extraction on raw seismic data, creating an approximate subsurface model early on. GeoDNN is not designed to replace the geoscientists (heaven forbid!) Other ML techniques help in well design, drilling and logging. GeoDNN’s geophysical feature detection is set to ‘greatly accelerate seismic processing’ but it is not (yet) perfect. GeoDNN was developed with help from MIT.
Shell is also using ML in reservoir engineering with a prototype tool for simulator post processing. Here the ‘AutoSum’ tool provides automatic summaries across large ensembles of reservoir models to help with understanding key sensitivities, ‘current tools are not adapted to this.’ AutoSum combines traditional physics-based models with analytics to identify the main contributions to geological uncertainty and minimize development strategy risk. If more data is required, the simulator can be run again to further investigate the uncertainty space. Machine learning has been applied to relate subsurface features with overall production strategy. Shell’s preferred ML tool is Python. Challenges remain. It is hard to persuade those used to physics-based models to switch. Data access and scale are issues as is the lack of ground truth (especially in seismics). But the approach has spin-off benefits. Data quality issues are found faster and the approach ‘frees up experts to focus on the important stuff.’ Finally, Limbeck warned of the crucial need to maintain underlying databases. [EarthDoc 89275]
Nicolas Audebert from the French R&D Onera establishment described the use of deep learning on hyperspectral data. The work was supported by Total as a part of its Naomi, ‘new advanced observation method integration*.’ Hyperspectral data comes from airborne and satellite-mounted sensors. The term refers to the wider than visual bandwidth, from infrared to UV, with a spatial resolution of around a meter. Imagery is used for geology and surface material identification. Convolutional neural nets have proved powerful for multimedia and RGB imagery. The neural net approach has been used since the 1970s on a pixel by pixel basis. Today, 2D/3D techniques use the full raw images, noise and all. Deep learning on these huge datasets has ‘blown everything else away.’ ADAM stochastic gradient descent, PyTorch and Nvidia Titan X GPUs also ran as did the Pavia Center dataset. Audebert is still on the lookout for more annotated data. ‘The potential of deep learning not yet fully realized.’ [EarthDoc 89272]
Henri Blondelle (Agile Data Decisions) outlined his work on the CDA unstructured data challenge. CDA provided a ‘fantastic dataset’ of logs and reports from decades of North Sea exploration. But much (80%) comes as scanned TIFF imagery or PDF documents. Blondelle wryly observed that the first task of a geoscientist is, ‘unstructure your data by making a print.’ The big question for the CDA challengers is how to put the structure back in. AgileDD’s IQC tool was used to classify documents, extracting metadata such as well status from reports, along with a measure of the probability of correctness. A range of tools contributed to the initiative, the Python ML library, Java, MySQL, JBoss Tools, Azure and a ‘hybrid cloud.’ CDA’s CS8 hardcopy catalogue formed the basis of a classification taxonomy, along with 2,000 annotated documents used as a training set. [EarthDoc 89273]
* Cheekily described in Le Figaro as a relaunch of the sniffer plane (avion renifleur).
© Oil IT Journal - all rights reserved.