CARST: A NEW CARBONATE STRATIGRAPHIC MODEL WITH INTEGRATED MACHINE LEARNING
All of the existing SFMs are forward models only. They cannot easily incorporate known data or knowledge. Instead, exploration of uncertainties in parameters is ether done by hand or by running thousands of potentially expensive simulations with the aim to minimise the difference in output to known data. Moreover, they use traditional software engineering technologies, which require extensive knowledge of both the geological processes, programming techniques, and computer hardware.
Since many of these models were created, new numerical modelling technology has been developed. Instead of writing traditional code (e.g. Fortran, C), a higher-level mathematical abstraction (i.e. the code reads as equations rather traditional loops or functions) is used which is then compiled into low-level code depending on the computer hardware being used. This methodology is a step-change in how numerical models are written: not only can models be written more quickly and code compiled for a target hardware (e.g. GPUs vs CPUs), but because the mathematical expressions are explicitly written, they can be differentiated symbolically. This creates new approaches in carbonate modelling using machine learning to ‘guide’ simulations.
Here, I describe the use of an adjoint solver, which calculates the gradient of a forward model in parameter space and as such can be used to assimilate data into a forward model run and to quantitatively explore parameter space. In other words, it can effectively invert the stratigraphic model by adjusting parameters to optimise a match to known data such as seismic and core data automatically. I show how this model can be used to simulate simple carbonate facies and how the adjoint solver can be used to constrain the parameter space and evaluate sensitivity of the model to some of the parameters.