Machine learning in geoscience

In this project we have developed EarthNET which is a cloud-native platform built to empower geoscientists with AI technology. EarthNET provides data-driven decisions support to geoscientists, reservoir engineers and E&P decision makers, enabling them to better understand the subsurface and hence to create and update sub-surface models with increased precision and efficiency. Petrophysicists, geophysicists, geologists, seismic interpreters, and E&P generalists can easily leverage advanced AI and data-science technology to maximize the value of their data, and build strong foundations for making data-driven decisions. EarthNET connects energy company users with their internal and external data assets, with high-performance computer resources, and with AI-powered geoscience software applications. This connectivity and the integrated applications allow EarthNET users to break out of the data and discipline silos and embrace a truly integrated and cross-disciplinary data analytics workflows. EarthNET is a platform that provides an integrated set of AI-powered tools. This integrated system allows geoscientists to access and use all the relevant data to gain new insights, to produce 3D property models and to compute data-driven probabilistic volumes for prospects and fields. The main purpose for EarthNET is to provide data-driven decisions support to subsurface specialists and decision makers. Our key applications are: EarthBANK: EarthNET gives users access to data through EarthBANK?s structured ML-Ready Databases. These databases include ?Data-Packs? provided by Earth Science Analytics, your energy company?s databases, and ?Data-Packs? from 3rd-party data providers. EarthAI Wells: With EarthNET?s well-data analytics module geoscientists can rapidly predict high-quality rock- and fluid-property curves and predict missing logs in wells. EarthAI well-to-seismic integration: EarthNET provides functionality that enable geoscientists to propagate knowledge from core to seismic scale. After predicting reservoir properties at the well scale, the next step is to expand our predictions to 3D. EarthAI 3D property prediction: With the 3D property prediction module geoscientists can rapidly predict high-quality rock- and fluid-property cubes, either as a function of inverted seismic data, or as a function of partial-stack cubes. Partial stacks are used as the features of the models, and labels are obtained from property curves in wells. EarthAI seismic interpretation: Automatic Seismic Interpretation (ASI) can speed up seismic interpretation workflows by a factor of >10. Geoscientists can easily train and/or apply machine-learning models for automatic seismic interpretation. Workflows that used to take weeks or months can now be done in days, while at the same time the quality of the interpretation is improved. The user can use its own labels, or apply one of the pre-trained ASI models available in EarthNET.

Resultatene fra PM2-prosjektet viser at det er mulig å anvende de enorme datamengdene som lenge har ligget ubrukt med ML-teknologi. Per i dag kan teknologien gjøre mange av oppgavene vi så for oss ved prosjektstart. Vi kan blant annet autotracke krevende horisonter og forkastninger som var utenkelig bare for noen år siden. Missed pay project Vi har gjennomført brønn analysere av over 8 000 brønner på norsk sokkel med vår ML-teknologi. Teknologien identifiserte over 200 intervaller med olje som mennesker ikke klarte å identifisere, altså geologen klarte ikke anvende de store datamengdene. Studiet ble også utført på relativt kort tid ettersom det krevde minimalt med menneskelig interaksjon. Tidligere måtte geologen blant annet gjette på mineraltetthet for hver seksjon som skulle analyseres.

Today a geoscientists spend 70 % of his or her time on seismic interpretation. Our proposed technology will, in short, allow automatic structural, stratigraphic and lithologic interpretation from seismic data in order to determine important geological formations such as faults, horizons, and geobodies such as salt domes, channels and other lithostratigraphic units. Furthermore, by utilizing deep neural networks we will be able to automatically infer rock and fluid properties from partial-stack and pre-stack seismic, which traditionally are extracted from cumbersome inversion techniques. We will combine the new proposed technology for seismic interpretation with EarthNETs current rock- and fluid property analytics from well data to provide accurate data driven decision support for petroleum exploration. Our approach to geoscience will in contrast to traditional methods be able to utilize the entirety of the available raw data that has been gathered since the beginning. This will not only dramatically increase the accuracy of the analysis, but do so in a fraction of the time compared to the traditional prospecting and exploration process. Our approach to geoscience represents a new paradigm for the entire industry and has the potential to revolutionize how hydrocarbon exploration is performed.