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

TL;DR

This paper introduces SCAN-ML, a machine learning model trained on ab initio data, enabling accurate molecular dynamics simulations of calcium carbonate formation from aqueous solutions, including reactive pathways.

Contribution

The paper presents a novel machine learning potential, SCAN-ML, that accurately reproduces ab initio results and captures chemical reactions in calcium carbonate formation.

Findings

SCAN-ML surpasses traditional force fields in accuracy

Reveals calcium bicarbonate binding as key in ion pairing

Enables study of reactive crystallization pathways

Abstract

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 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 semi-empirical 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 study of reactive crystallization pathways in calcium carbonate, which are currently poorly understood.