Exploring the polymorphism of dicalcium silicates using transfer learning enhanced machine learning atomic potentials

Belitic cements are a greener alternative to Ordinary Portland Cements due to the lower CO2 associated to their production. However, their low reactivity with water is currently a drawback, resulting in longer setting times. In this study, we utilize a combination of evolutionary algorithms and machine learning atomic potentials (MLPs) to identify previously unreported belite polymorphs that may exhibit higher hydraulic reactivity than the known phases. To address the high computational demand of this methodology, we propose a novel transfer learning approach for generating MLPs. First, the models are pre-trained on a large set of classical data (ReaxFF) and then re-trained with Density Functional Theory (DFT) data. We demonstrate that the transfer learning enhanced potentials exhibit higher accuracy, require less training data, and are more transferable than those trained exclusively on DFT data. The generated machine learning potential enables a fast, exhaustive, and reliable exploration of the dicalcium silicate polymorphs. This includes studying their stability through phonon analysis and calculating their structural and elastic properties. Overall, we identify ten new belite polymorphs within the energy range of the existing ones, including a layered phase with potentially high reactivity.