PHD PROJECT: UNCERTAINTY QUANTIFICATION OF GEOMECHANICALLY SENSITIVE RESERVOIRS USING PHYSICS-INFUSED MACHINE LEARNING
What is the problem?
When we use the subsurface to store CO2 or Hydrogen, the rocks can deform inelastically as the reservoir conditions change – processes well described in geomechanics. But this permanent strain can lead to significant leakage and to induced seismicity risks during operation and/or long term storage. At worst, these risks pose a substantial threat to property and health – so much so that regulators and investors may halt projects before they start as we can’t correctly estimate the geomechanically-derived risks.
The key technical challenge for most storage reservoirs is to accurately quantify uncertainties and risks associated with geomechanical property changes from a few computer model forecasts. Geomechanical simulation software is complex and data-intensive, each run taking many hours to days. Such long run times make statistically thorough methods to quantifying uncertainty impractical. In most circumstances, we can’t afford the thousands of required model realisations. So such risks may be misestimated or even missed.
The longest model run times occur when we couple simulations of fluid flow (production and injection) with geomechanical simulations to predict how a development plan may alter the reservoir rocks properties and how this will impact fluid movement and field operations. To solve the system of equations for fluid flow and geomechanics together, we need to connect very different modelling approaches, the differences in the solvers typically precluding full coupling. Instead, the packages interact by running separately but simultaneously and passing data back and forth between iterations. This is a technically monumental challenge and the run times of the models are far longer than the combined times of each model run separately.
To accurately quantify uncertainties in geomechanically sensitive reservoirs we must run many more models than we do today, exploring a more diverse set of geological scenarios. But to run more models, we must significantly improve the efficiency of coupling fluid flow and geomechanical simulations. Machine learning provides one solution: once trained on an appropriate data set it can capture complex, non-linear systems very rapidly.
Continue reading →