One of the big questions in enhanced weathering and its applications revolves around the formation of secondary minerals—those that form when the rock dissolves. This affects carbon sequestration potential and efficiency. Direct observation of those secondary minerals is difficult, if not impossible, because of the small relative amounts with respect to the main rock mass. But also because they presumably first form with disordered structures or as amorphous precursors, which prohibits detection with X-ray diffraction methods.
In the BAM project we approach this as a reverse problem – we try to predict the most likely combination of secondary minerals that are forming, based on the detailed observation in the water that circulates through the rock bed. For this, we combine geochemical modeling with machine learning.
The geochemical model simulates the interactions between water, CO₂, and minerals inside the reactor, and computes the concentration in the water phase that are measured continuously. The mismatch between observation and simulation can in theory tell us whether all chemical processes are accounted for or not. But there’s a catch: many of these processes are difficult to quantify directly. For example, as minerals dissolve, the reactive surface area—the part of the rock that interacts with water—decreases, a phenomenon for which we only have very approximative models. Here’s where machine learning comes in. By training, we can estimate the decline in reactive surface area more accurately, and untangle this process from the other chemical and physical processes that occur inside the rock bed.
This hybrid method lets us refine our understanding of the system dynamically, improving predictions of CO₂ uptake and mineral transformation over time. The goal? A smarter, more transparent reactor design that maximizes carbon drawdown while minimizing surprises. By making the invisible chemistry inside these reactors more visible, we’re a step closer to scalable, effective climate solutions.