Marko Maucec
PE Consultant
Saudi Aramco
Participates in
TECHNICAL PROGRAMME | Energy Technologies
The Energy Transition: The Role of Digitalisation, AI, and Cybersecurity
Forum 23 | Digital Poster Plaza 4
30
April
10:00
12:00
UTC+3
Objectives/Scope:
We have previously shown that deep interaction neural networks based on graphs, can learn complex flow physics relationships from reservoir models to accelerate forward simulations [1]. The generalization and scale-up to a deep learning architecture with Subsurface Graph Network Simulator (SGNS) [2] renders fast and accurate, long-term spatio-temporal predictions of fluid and pressure propagation in structurally diverse reservoir model grids [3]. This paper builds on the later and introduces the benchmarking and deployment of SGNS in operational engineering simulation environment.
Methods, Procedures, Process:
The SGNS is a data-driven surrogate framework that consists of a subsurface graph neural network (SGNN) to model the evolution of fluids, and a 3D-U-Net to model the evolution of pressure. The SGNN uses encoder-processor-decoder architecture to encode the properties and dynamics of each grid cell and the cell-cell relation into graph node and edge features. The multilayer perceptron computes the interaction between neighboring cells and updates the state of the cells. The pressure dynamics manifests shorter equilibrium time and faster global spatial evolution than the fluid and is better captured with hierarchical structure and modified order of operation of 3D-U-Net convolution layers.
Results, Observations, Conclusions:
We deploy the SGNS on a synthetic, single-porosity/single-permeability reservoir model with multi-phase flow, large-scale simulation grid and large numbers of injectors and producers with variable positioning and geometry. We construct a network graph with nodes, representing reservoir grid cells are encoded with tens of static, dynamic, computed and control features. The wells are encoded via well completion factors. The graph edges represent interactions between the nodes with encoded features like transmissibility, direction and fluxes. We implement sector-based training with multi-step rollout to avail for the use of large-scale models. The loss function is the joint mean squared error, combining misfits in oil and water volumes and pressure. We present the comparative results between the SGNS and the full-physics simulation for up to 30-year prediction of the 3D pressure, oil and water saturation as well as the dynamic well responses.
Novelty/Significance/Additive Information:
The SGNS framework is a novel, industry-unique technology with promising scalability, generalization and prediction accuracy. The immediate applications involve accelerated well placement and production forecasting studies. Going forward, we are in the process of integrating the well prediction model in the roll-out of the evolution model, incorporating the encoding of well production constraints into the training scheme and deploying state-of-the-art architectures to enable multi-GPU training.
References:
We have previously shown that deep interaction neural networks based on graphs, can learn complex flow physics relationships from reservoir models to accelerate forward simulations [1]. The generalization and scale-up to a deep learning architecture with Subsurface Graph Network Simulator (SGNS) [2] renders fast and accurate, long-term spatio-temporal predictions of fluid and pressure propagation in structurally diverse reservoir model grids [3]. This paper builds on the later and introduces the benchmarking and deployment of SGNS in operational engineering simulation environment.
Methods, Procedures, Process:
The SGNS is a data-driven surrogate framework that consists of a subsurface graph neural network (SGNN) to model the evolution of fluids, and a 3D-U-Net to model the evolution of pressure. The SGNN uses encoder-processor-decoder architecture to encode the properties and dynamics of each grid cell and the cell-cell relation into graph node and edge features. The multilayer perceptron computes the interaction between neighboring cells and updates the state of the cells. The pressure dynamics manifests shorter equilibrium time and faster global spatial evolution than the fluid and is better captured with hierarchical structure and modified order of operation of 3D-U-Net convolution layers.
Results, Observations, Conclusions:
We deploy the SGNS on a synthetic, single-porosity/single-permeability reservoir model with multi-phase flow, large-scale simulation grid and large numbers of injectors and producers with variable positioning and geometry. We construct a network graph with nodes, representing reservoir grid cells are encoded with tens of static, dynamic, computed and control features. The wells are encoded via well completion factors. The graph edges represent interactions between the nodes with encoded features like transmissibility, direction and fluxes. We implement sector-based training with multi-step rollout to avail for the use of large-scale models. The loss function is the joint mean squared error, combining misfits in oil and water volumes and pressure. We present the comparative results between the SGNS and the full-physics simulation for up to 30-year prediction of the 3D pressure, oil and water saturation as well as the dynamic well responses.
Novelty/Significance/Additive Information:
The SGNS framework is a novel, industry-unique technology with promising scalability, generalization and prediction accuracy. The immediate applications involve accelerated well placement and production forecasting studies. Going forward, we are in the process of integrating the well prediction model in the roll-out of the evolution model, incorporating the encoding of well production constraints into the training scheme and deploying state-of-the-art architectures to enable multi-GPU training.
References:
- M. Maucec, R. Jalali (2022). “GeoDIN-Geoscience-Based Deep Interaction Networks for Predicting Flow Dynamics in Reservoir Simulation Models”, SPE-203952-PA, SPE Journal, 27 (03): 1671–1689
- T. Wu et al. (2022). “Learning Large-scale Subsurface Simulations with a Hybrid Graph Network Simulator”, 2022 ACM SIGKDD.
- M. Maucec, R. Jalali; H. Hamam (2024). “Predicting Subsurface Reservoir Flow Dynamics at Scale with Hybrid Neural Network Simulator”. IPTC-24367-MS.
