Overview
Accurately predicting subsurface behavior is essential for many energy transition technologies, from geothermal energy and carbon capture and storage (CCS) to hydrogen and energy storage. Modelling geological processes through geostatistics or forward modelling provides the foundation for reliable predictions of how the subsurface evolves over time. Geostatistical approaches help capture spatial variability in key rock properties, enabling better predictions of reservoir performance and uncertainty quantification. On the other hand, forward process-based modelling simulates how geological systems evolve, driven by fundamental processes such as sedimentation, diagenesis, fluid flow, and heat transfer. By integrating these two approaches, geoscientists can better anticipate how reservoirs respond to thermal and pressure changes over time, ensuring the safe and efficient deployment of energy transition technologies. These models are vital for minimizing risks, optimizing storage capacity, and enhancing the sustainability of subsurface energy solutions in a rapidly changing world.
Forward Stratigraphic Models
In our research, we combine field methods to gather information on the sediment and stratigraphy of carbonates with a numerical technique called “forward stratigraphic modeling”. The principle is to model forward in time how sediments within a basin are being deposited, eroded and transported, thus allowing to predict the nature of the sediment and the stratigraphic architecture of the rocks. Importantly, uncertainties in the stratigraphic models can be determined by repeating the modeling numerous times, allowing the establishement of uncertainty maps.
Forward Stratigraphic Modelling of the Lower Cretaceous (Oman). This example shows the use of forward modelling to simulate the deposition of sediments of Cretaceous age in Oman. We can use this numerical laboratory to, for instance, estimate subsidence rates or sediment supply, and use the known stacking pattern of the rock record as our control experiment.
Typically, we define a region to model (a basin) and a time in the geological past, and from there, we can model forward how sediments are being deposited, eroded and transported in the basin, where the main area of deposition are, what the nature of the sediment deposited is, and the nature of the rocks in the subsurface. Importantly, we can also derive the uncertainties in our models, e.g. build uncertainty maps. We use diffusion equations to represent sediment flow, an approach that has been shown to work well for sediments modeled over long (>10’000 years) periods.
Examples of problems we solved using this approach include understanding the regional 4D (three dimensional, plus geological time) distribution of carbonate sediments in the Early Cretaceous of Oman, a time periods characterised in the Middle East by the deposition of important hydrocarbon reservoirs. This study is now published in the Journal of Marine and Petroleum Geology (see Al-Salmi et al, 2019), and we demonstrate for instance the need for regional differential tectonic in the Early Cretaceous to explain the stacking patterns of carbonate rocks in this area.
Other examples include understanding the sub-salt of Kazakhstan, where we looked at the Serpukovian Stage of the Karachaganak FIeld (John et al, 2021). There, we worked on a relatively small (10-20 km) carbonate atoll, and tried to understand the constraints from the data on the geologic models. We demonstrated for instance that the seismic horizon picking needed to be re-evaluated in view of our models, and we also were able to model where and why the most promising reservoir facies would be.
A final example is our work on organic-rich shales of the Najma Formation in the subsurface of Kuwait, where we explore how much oxygen must have existed at the bottom of the shallow sea to account for the amount of organic material preserved in the sediments (Al-Wazzan et al, 2022).
Geostatistics and other numerical approaches
Numerical methods in geosciences are not limited to forward stratigraphic modelling. A very common approach is known as “geostatistics”. Traditionally, geostatics refers to the Krigging method, where the geospatial distribution of a property of interest can be modelled using chacateristics of a distance-based variogram. In our work, we developed pluri-gaussian simulations to model carbonate rocks at a range of scales. This work, writen using the programming language ‘R’, has resulted in a number of publications (Le Blevec et al, 2017, 2020 – example result below).
Other examples of numerical work within our research group include fluid flow modelling with experimental design to model the effect of various geologic heterogeneities within a carbonate reservoir on production scenarios (Fitch et al, 2014), and surface-based modelling of Jurassic carbonates from Saudi Arabia (Jacquemyn et al, 2018).
The Role of Artificial Intelligence in Process Modelling
You might wonder what the role of artificial intelligence and deep learning is in modelling the subsurface. One disadvantage of forward models is that they can be very computationally demanding. This limits the number of realisation that can be achieved, as the simulation time becomes very long. By extension, this means that capturing modelling uncertainties becomes difficult. Deep learning can overcome this barrier by offering simulation speeds oder of magnitude faster than process-based models.
For instance, we have explored how deep generative artificial intelligence can be leveraged to generate ‘steps’ in a forward-process framework. This is akin to geostatistics, where we sample the data distribution of a process model with the conditioning of the state of the previous step, to generate the next step. In her MSc work, Adiba Feizal (2023) was able to generate realistic looking plan views of an isolated carbonate platform using Generative Adversarial Networks (GANs – see image below). The beauty of this approach is that it takes a fraction of a second to generate, rather than 15 minutes of computations.
We currently are leveraging this experience, and working on an innovative project that will combine data assimilation and deep learning surrogate models, two technologies readily adopted by weather and climate model. By exploring their integration into geologic forward models, our aim is to accelerate simulations speed and enhances calibration for subsurface energy applications. Looking forward, we see the potential of digital twins for the subsurface, enabling real-time simulations to optimize the prospection and production of clean energy such as for instance in geothermal fields.
