Introduction
FlowNet aims at making predictive subsurface models directly from data. Instead of complex simulation grids, FlowNet uses networks of tubes (with full physics within each tube), speeding up simulation models by order of magnitudes, which again makes it possible with e.g. increased sampling of the uncertainty span compared with traditional workflows.
Key features
Fast model building: Very efficiently get an ensemble of subsurface models, compared to duration of traditional modeling projects.
Fast simulation runtime: Reduce the simulation time per model realization by order of magnitudes, which can be traded for e.g. better sampling of your uncertainty span.
Automatic parametrization: FlowNet will automatically parametrize typical subsurface uncertainties for you.
Data driven: FlowNet will use what typically is considered raw data in a model building process directly.
Easy: As FlowNet user, you only need to adjust a high level configuration file - no Python skills necessary.
Standards-based: Based on and compatible with open standards. FlowNets stands on the shoulders of giants:
ERT for doing uncertinaty sampling (and optionally assisted history matching).
Flow for running the simulation models.
Webviz for visualization of results (profiles, history match quality, parameter distributions…)
ResInsight for three dimensional exploration of the network and simulation grid results.
Overview
FlowNet aims at being data driven software. I.e. the preferred way is that raw input data to FlowNet should arrive from APIs (from other softwares). However, in order to also facilitate testing and developing FlowNet without an available data API framework, FlowNet can also extract necessary input data from an existing simulation model.
Automatic parametrization
In traditional modelling workflows, the user has to parametrize the model manually when setting it up. This is time consuming and routine work, and can also result in less parametrization (which again might result in not spanning the real uncertainty span).
FlowNet will automatically add typical parametrizations, and leave only responsibility for high level options (e.g. prior-distribution end-points) to the user, which he/she sets through the configuration file.
FlowNet currently supports the following automatic parametrization:
Aquifers
Equilibration
Fault transmissibilities
Permeabilities
Porosity and porevolumes
Relative permeabilities
Rock compressibility
To be added:
PVT table parametrization
Lift curve parametrization
…
Benchmark models
The core developers of FlowNet are currently testing on the following open benchmark models:
A small Norne assisted history matching example using FlowNet is automatically started on each commit to FlowNet pull requests, in order verify the code changes.
Going forward, we will try to aim for having an Azure resource running on merge to master, and then have the results (profiles, history match profiles etc.) automatically pushed back to the FlowNet repository - showing current results on the benchmark model.
Publications
A first version FlowNet was presented at the 17th European Conference on the Mathematics of Oil Recovery (ECMOR XVII) held 14-17 September 2020. The reference for the conference paper is
A.F. Kiær, O.P. Lødøen, W.J. de Bruin, E. Barros and O. Leeuwenburgh, Evaluation of a Data-Driven Flow Network Model (FlowNet) for Reservoir Prediction and Optimzation, Conference Proceedings, ECMOR XVII, Sep 2020, Volume 2020, p.1-18, DOI: https://doi.org/10.3997/2214-4609.202035099