Towards digital companions for subsurface modellers: combining data science and natural language processing to facilitate geomodelling workflows

Antoine Bouziat and François Cokelaer and Renaud Divies and Sylvain Desroziers and Mathieu Feraille and Jeanneth Bouziat. ( 2020 )
in: 2020 RING Meeting, ASGA

Abstract

With the commercial success of Amazon Alexa or Apple Siri, the term “digital companion” has been popularised beyond scientific and technological circles to reach the general public. It usually refers to devices based on machine learning and natural language processing (NLP) approaches, designed to assist their owners in various routine tasks. In this study, we explore the extension of this concept to the geomodelling domain. We appraise the potential of recent data science and NLP technologies to accelerate and democratise several steps of the geomodelling routine, from mining literature for parameters to browsing physical simulation results. We notably illustrate their value focusing on basin modelling in a petroleum exploration context. Firstly, we assess text entity extraction techniques to automatically collect information about hydrocarbon source rocks in scientific papers. We build a dedicated ontology representing the conceptual relationships between source rock features, and we manually annotate 127 papers accordingly. Then we train an industrial deep learning model to autonomously retrieve similar features from a larger bibliographic corpus. The pieces of information extracted and their relationships are stored in a specific graph data base. Eventually, to ease the graph exploration, we build an interpreter of natural language queries and interface it in a web application. As a result, operational geoscientists can select modelling hypotheses and discuss simulation results from a wide literature in an efficient and intuitive fashion. Secondly, we evaluate innovative ways to facilitate the analysis of physical simulation results. We start with linking a 4D basin model with interactive data visualisation dashboards specifically tailored to highlight the maturity evolution of the source rocks through their geological history. Then we further democratise the process by training a full conversational engine to interpret natural language queries, to browse the simulation results for an answer and to provide a relevant data visualisation. The resulting tool relies on an industrial and fully interfaced cloud-based platform. At the end, users from diverse backgrounds can interact with the geomodel as fluently as they would chat with their friends or colleagues. We consider these examples pave the way for fully integrated earth-modelling companions, which could in the future assist many professionals in their geomodelling work.

Download / Links

BibTeX Reference

@INPROCEEDINGS{BOUZIAT_RM2020_2,
    author = { Bouziat, Antoine and Cokelaer, François and Divies, Renaud and Desroziers, Sylvain and Feraille, Mathieu and Bouziat, Jeanneth },
     title = { Towards digital companions for subsurface modellers: combining data science and natural language processing to facilitate geomodelling workflows },
 booktitle = { 2020 RING Meeting },
      year = { 2020 },
 publisher = { ASGA },
  abstract = { With the commercial success of Amazon Alexa or Apple Siri, the term “digital companion” has been popularised beyond scientific and technological circles to reach the general public. It usually refers to devices based on machine learning and natural language processing (NLP) approaches, designed to assist their owners in various routine tasks. In this study, we explore the extension of this concept to the geomodelling domain. We appraise the potential of recent data science and NLP technologies to accelerate and democratise several steps of the geomodelling routine, from mining literature for parameters to browsing physical simulation results. We notably illustrate their value focusing on basin modelling in a petroleum exploration context. Firstly, we assess text entity extraction techniques to automatically collect information about hydrocarbon source rocks in scientific papers. We build a dedicated ontology representing the conceptual relationships between source rock features, and we manually annotate 127 papers accordingly. Then we train an industrial deep learning model to autonomously retrieve similar features from a larger bibliographic corpus. The pieces of information extracted and their relationships are stored in a specific graph data base. Eventually, to ease the graph exploration, we build an interpreter of natural language queries and interface it in a web application. As a result, operational geoscientists can select modelling hypotheses and discuss simulation results from a wide literature in an efficient and intuitive fashion. Secondly, we evaluate innovative ways to facilitate the analysis of physical simulation results. We start with linking a 4D basin model with interactive data visualisation dashboards specifically tailored to highlight the maturity evolution of the source rocks through their geological history. Then we further democratise the process by training a full conversational engine to interpret natural language queries, to browse the simulation results for an answer and to provide a relevant data visualisation. The resulting tool relies on an industrial and fully interfaced cloud-based platform. At the end, users from diverse backgrounds can interact with the geomodel as fluently as they would chat with their friends or colleagues. We consider these examples pave the way for fully integrated earth-modelling companions, which could in the future assist many professionals in their geomodelling work. }
}