A recent French Geological Society gathering looked into the future of geoscience jobs in the light of the rise of machine learning and artificial intelligence. Total’s Yves Le Stunff presented on progress in AI and digital technologies in geoscience. Le Stunff described the two contrasting scientific approaches, on the one hand, physical theory and numerical modeling, on the other hand, data driven discovery and machine learning (ML). ML is not new, Total has been using it for twenty years, with a lot of work in the 1990s on classification. Today, computers are more powerful and there is lot more data. The work done in the 1990s is now being revisited across the upstream, in support of technology watch, through E&P to production. Total is currently investigating the application of AI in kick prediction, stuck pipe and predictive maintenance. In the geoscience area, AI is being investigated for automatic classification of micro paleontology in thin sections and on satellite image tracking for seeps and spills. In the field of unconventionals, the physics is not very well understood, but there is a lot of data where AI is being tried to predict production decline. Total has initiated the Gaia program (with Google) to test AI in geoscience.
What will be the likely impact of all this? Will AI prove a game changer in oil and gas? Today Total has some 600 geoscience interpreters – so these technologies ‘could have a significant impact’. Today, geoscientists spend time looking for and manipulating information. The knowledge work is often constrained by a deadline. Gaia’s objective is to give time back to the geoscientist with a ‘virtual assistant’ (like Apple’s Siri) that screens data and speeds-up interpretation by proposing a solution to the user.
But the ML part is not as hard to realize as the data collection and ‘productizing’, which requires a lot of IT know-how. The Gaia results will be delivered as Sismage* plug ins. Le Stunff cautioned that ML is not magic and doesn’t work all the time. If data is clean, results can be remarkable, saving up to 80% of the time in fault mapping. If the data is dirty, things are not so great! It may take as much time to clean the data as it does to perform the ML work. Overall, average fault pick productivity is probably around 30%. For stratigraphic interpretation (horizons and geobodies) the results are interesting but not usable without rework. Interpreter-assisted interpretation is better, but again, noise levels are critical results are poor in areas of complex geology.
Another line of research is the ‘semantic stream’ i.e. natural language processing where Google excels. This is intended to speed document and information retrieval from reports, from Wikipedia and other sources. Total is developing a web app that goes beyond the search engine, enabling object detection in documents with named entities recognition (well, fields) and classification of maps, cross plots and seismics. This is all at the PoC stage currently and the outcome is uncertain. It is a harder problem than seismics. Some components already work – like image classification with Google AutoML and can recognize for example a geological map. Extracting tables from a PDF document is ‘emerging’. Total is building a knowledge graph to support better natural language query. Initially, the expectation was that algorithm development would take time. In fact, Total spent more time thinking about what to do with ML, defining KPIs and building infrastructure. Integration is much more time consuming that the optimization. ML is ‘never 100% some projects never stop’.
* Sismage is Total’s in-house developed seismic interpretation environment.
Steve Purves gave a compelling presentation on the technology that Norwegian Earth Analytics has developed to apply AI across the upstream workflow. Studies using Norwegian Petroleum Directorate (NPD) data have revealed ‘bias and errors’ in human-driven process. More sophisticated ways of analyzing geoscience data are needed, using computer vision, NLP and graph databases. There is a lot of data out there, but it is not clean or convenient and is a struggle to use. The AI/ML application landscape is balkanized, with pockets of individual apps. EarthAnalytics provides a stack of a structured, ML-ready database (EarthBank) and EarthNet a suite of AI-based interpretation tools. The Norwegian EarthBank is a cleansed edition of the national Diskos dataset. EA is currently working to produce an Earthbank for the UK from OGA data. The combo allows a petrophysicist to analyze hundreds of wells in an afternoon or a seismic interpreter to perform rock physics-based seismic inversion in hours instead of months. The technology was developed with a combination of manual labelling on 3D seismic and (increasingly) synthetic data from models. EA sees potential for data-driven pipelines to capture the accuracy and uncertainty of models from hundreds of realizations. ‘A new wave of ML is already disrupting subsurface workflows. With increasing automation, data availability and better data management, ML adoption will continue to accelerate’.
Cynthia Gomez presented Seisnetics’ ML for seismic processing and interpretation. Seisnetics uses a database of ‘100s of millions of seismic waveforms’ obtained from unsupervised ML to speed up and augment seismic interpretation and processing. The technology applies learnings ‘from the human genome project’, with the seismic waveform replacing the chromosome along with ‘natural selection and survival of the fittest’. Waveforms are grouped into a ‘geopopulation’, of genetically and spatially related types. The system is 100% data-driven, learning from the data. A test on the SEG SEAM dataset correctly extracted the SEG logo! Sub seismic resolution is claimed.
Meeting Chair Henri Blondelle also presented AgileDD’s work on extracting meaningful information from legacy documents and reports with Yolo which we cover in our report from ECIM also this issue.
© Oil IT Journal - all rights reserved.