Abstract
Machine learning potential is an emerging and powerful approach with which to address the challenges of achieving both accuracy and efficiency in molecular dynamics simulations. However, the development of machine learning potentials necessitates intricate construction of descriptors, particularly for complex material systems. Therefore, the Deep Potential method, which utilizes artificial neural networks to autonomously construct descriptors, are employed to develop a deep learning-based potential for calcium silicate hydrates (the basic building block of cement-based materials) in this study. The accuracy of this potential is validated through calculations of energetics, structural, and elastic properties, demonstrating alignment with first principle calculations and an efficiency 2–3 orders of magnitude higher. Additionally, the deep potential successfully reproduces precise predictions in C-S-H models with different calcium-to-silicon ratios, thereby confirming its remarkable transferability. This potential is expected to fulfill cross-scale computations and bottom-up design of cement-based materials with both high accuracy and efficiency.
Disclosure statement
No potential conflict of interest was reported by the authors.
Data availability
The compressed DP model for C-S-H generated in this work is available in the Figshare: https://doi.org/10.6084/m9.figshare.19608555. And the database of used is available in the Figshare: https://doi.org/10.6084/m9.figshare.19608549.