Machine Learning Techniques for Chemical Prediction during Carbon Capture and Storage

Carbon capture and storage (CCS) is an essential part of limiting anthropogenic climate change. The carbon storage aspect of CCS involves the injection of CO2-rich fluid deep underground, typically in abandoned oil reservoirs. Once injected, the CO2-rich fluid can interact with the rocks in a number of ways. Some of these are positive, including the mineralisation of carbon which prevents leakage. However, some of these are negative, including those which drive changes in pH and lead to mineral dissolution or extreme mineral precipitation, influencing secondary porosity evolution. Fluid chemistry in various natural environments in particular can be amended through the action of microbial communities in the environment. In particular, microbial sulfate reduction (MSR) and other redox-sensitive microbial metabolisms can drastically alter fluid chemistry. This metabolism influences pH and generates dissolved inorganic carbon (DIC), which in turn changes the saturation state of minerals within the system. Ultimately, these microbial metabolisms can have a large effect on the mineralogy and overall porosity and permeability structure of the geochemical system.

The student will use reinforcement machine-learning coupled to a geochemical reactive-transport model to predict the optimal fluid composition for CCS. Given a site for CCS in a geochemical state A, with a target geochemical state B (that might be more conducive to a successful/efficient storage of carbon) the model will predict the optimum conditions (in this case a series of geochemical amendments, defined in terms of chemical composition of amendment, duration of injection, location of injection etc.) needed. One of the advantages of the reinforcement learning approach is that the model, once trained, is agnostic as to the target state. Target conditions could be specified in terms of total mass of CO2 stored, total alkalinity, mineral volumes, or even isotope composition.

The reinforcement machine-learning will be trained using a multi-dimensional isotope-enabled reactive-transport model. As a first step we currently work with a neural network that can predict geochemical changes during CCS, given an initial state. This investigation is based on generating and interrogating thousands of unique geochemical scenarios and feeding them directly into a neural network for training and testing. Currently we have a regression model that can predict pH given a random set of initial conditions during CCS. There is scope for expansion of this technique to predict any geochemical quantity of interest. The student will develop this underlying, predictive model to train the reinforcement learning model to predict geochemical amendment policy, which will hopefully have industrial applications in CCS, but also any problem where secondary porosity evolution is of interest, as well any other application where we wish to develop strategies to effect a specific changes geochemical systems.