Ab initio machine-learning simulation of calcium carbonate from aqueous solutions to the solid state.

A first-principles machine-learning model has been developed aimed at studying the formation of calcium carbonate from aqueous solution using molecular dynamics simulations. The model, dubbed strongly constrained and appropriately normed-machine learning (SCAN-ML), reproduces accurately the potential energy surface derived from ab initio density-functional theory within the SCAN approximation for the exchange and correlation functional. A broad range of properties have been calculated relevant to ions in solution, solid phases, and the calcite/water interface. Careful comparison with results from experiments and semiempirical force fields shows that SCAN-ML provides an excellent description of this system, surpassing state-of-the-art force fields for many properties, while providing a benchmark for many quantities that are currently beyond the reach of direct ab initio molecular dynamics. A key feature of SCAN-ML is its ability to capture chemical reactions, which reveals that calcium carbonate ion pair formation occurs predominantly via binding of calcium to bicarbonate, with the subsequent loss of a proton to water, rather than by direct association. Our model thus paves the way for the ab initio study of reactive crystallization pathways in biominerals, which are currently poorly understood.