Large volumes of subsurface data exist, but current workflows for subsurface understanding are suboptimal, resulting in inadequate utilization of datasets.
Using an ensemble of model predictions to support robust decision-making is in its infancy; thus, we must establish consistent methods for robust decision-making. Digitalization and machine learning (ML) are required components of a sustainable subsurface value chain, and we must integrate knowledge and competence building to make more informed decisions. We will establish a digital infrastructure, i.e., Subsurface Knowledge Cloud (SKC), to provide readily usable data, high-performance computing power, and visualization tools. In work package 5:
- We facilitate robust predictions with feasible computational costs and big-data accessibility.
- We perform reliable uncertainty quantification for multi-purpose reservoir usage.
- We develop data-driven approaches to integrate ML into subsurface characterization, uncertainty quantification, and the decision-making process.
Work package 5 summed up
Large amounts of data are available, and because of advances in sensors and technology more and more data are constantly gathered and becoming available. In WP5 we develop digital tools which honor the measured data and are robust in predicting the future. The digital tools help us minimize the cost, maximize the profit and minimize the environmental footprint.
Six projects have been defined:
This project creates a Federated Knowledge Cloud that will serve as the cloud infrastructure and AI platform for subsurface digital integration in NCS2030. It aims to enable users to de-velop, deploy, and execute AI projects efficiently. Moreover, it brings together cloud services, federated learning, marketing, and assisting tools in enabling seamless across-silos collabo-rations for advanced knowledge gain, improved decisions, and efficient workflows.
Increasing model complexities and a desire to include multi-scenario models, including, e.g., various geological settings, make the problem of uncertainty quantification a daunting com-putational challenge. A compelling way of handling computational issues is to use a single, or multiple, computational models with reduced fidelity. Methodologies for robust probabilistic production forecasts, utilizing scenarios and fidelity models, will be developed.
In ensemble-based reservoir management, one typically runs many reservoir models in parallel. A reservoir-model workflow can include several software and scripts, and automation is essential. This project will provide the ensemble tools needed to automate the simulation of the ensemble of workflows and the ensemble updates through history matching and optimization.
Ensemble data assimilation (DA) and optimization methods are popular approaches to sub-surface characterization, development and management problems. Meanwhile, machine learning (ML) has emerged as a powerful toolset with a variety of applications in subsurface problems. The similarities and connections among DA, optimization and ML pave the way of developing advanced DA and optimization algorithms that are powered by modern ML tech-nologies, and have the potential to go beyond the current state-of-the-art.
Utilize Open Earth Community. Open Earth Community (OEC): Landmark Graphics to provide access to a full open development environment which includes all development tools and all Halliburton Landmark solutions. The Data Analytics platform can be utilized for building ML models. Pre-models are available that can be re-used or extended. In addition, Landmark Graphics can provide access to DISKOS data upon approval from NPD.
Schlumberger will assist NCS2030 in getting trained and utilizing modern OSDU-based workflows by getting access to the DELFI Cognitive E&P Environment. DELFI has a modern OpenAPI based Developer environment and a flexible Data science and Analytics platform leveraging Dataiku and TIBCO’s Spotfire. In collaboration with NCS2030 researchers, topics like Real-time Reservoir Optimization, Proxy Modelling and Data-driven Physics-based Predic-tive Modelling can be investigated.
