Postgraduate research opportunities Machine Learning for Forecasting Dynamics of Geological Reservoir during CO2 or H2 Storage

According to International Energy Agency’s 2021 world energy outlook report, a substantial amount of energy will be needed from fossil fuels in all scenarios. Therefore, it is essential to capture and store the CO2 in underground saline aquifers and move towards hydrogen economy.

Substantial storage capacity for CO2 and H2 is needed for which subsurface geological reservoirs are pertinent. To understand the capacity of an aquifer or old oil and gas reservoirs, and our ability to store gases in these, studying fluid dynamics in these geological reservoirs is essential.

We use numerical simulations because it is difficult to scale up analytical models with reservoir complexity. These physics-based reservoir simulations that we use to understand the flow and transport in the reservoirs are numerically very expensive due to the size of the reservoir. Therefore, upscaling of the geological properties such as porosity and permeability is required to attempt the numerical simulations for these reservoirs.

However, by upscaling we lose information on the heterogeneity of the system. Upscaling, or homogenization, is substituting a heterogeneous property region consisting of fine grid cells with an equivalent homogeneous region made up of a single coarse-grid cell with an effective property value. Equivalent, in this case, means either volume or flux must be the same in the fine-scale and upscaled model, depending on the type of property that is to be upscaled.

Upscaling is performed for each of the cells in the coarse grid and each of the grid properties needed in the reservoir flow-simulation model. Therefore, the upscaling process is essentially an averaging procedure in which the static properties (Eg. Porosity and permeability) and dynamic properties (Eg: saturation and Pressure) of a fine-scale model are to be approximated by that of a coarse-scale model.

The upscaling reduces the simulation time to efficiently predict the flow in a geological system. We use these simulations to understand various gas injection, storage and production scenarios, where we are essentially repeating a task for which we can have a proxy simulator after running only a few simulations for a few scenarios. Therefore, upscaling and running repeat simulations for various scenarios can be solved using machine learning.

Challenges

For the above two challenges we would like to answer the following questions:

1. Can we use super-resolution techniques, generally used for static images, to find out flow dynamics in the fine-scale model while we simulate in the upscale model of the geological reservoir? The super-resolution procedure that reconstructs the data such as saturation profiles and pressure profiles at fine scale needs to be physically, temporally and spatially consistent.
2. Can we have a proxy machine learning model for our reservoir simulator and learn the geology and flow behaviour using simulation results? The simulations in the upscaled model are still computationally expensive due to the size of the reservoir and time scale for which we need to predict the storage capacity and flow behaviour. Proxy models for forecasting such as polynomial fitting and design of experiments, for various injection strategies of CO2 and H2 and production of H2 do not scale well with the size and complexity of the simulations as we know from our experience in oil and gas reservoirs. We want to learn a Machine Learning proxy model to learn the complex dynamics of a reservoir and make a time-series prediction.