A high-efficiency neuroevolution potential for tobermorite and calcium silicate hydrate systems with ab initio accuracy

TL;DR

This paper introduces a highly efficient machine learning potential for tobermorite and C-S-H systems that achieves near-DFT accuracy with minimal training data and enables large-scale, GPU-accelerated molecular dynamics simulations.

Contribution

The authors develop a novel neuroevolution-based machine learning potential that combines high accuracy with computational efficiency for cement-related materials.

Findings

Achieves DFT-level accuracy with only 300 training structures.

Enables GPU-accelerated simulations of thousands of atoms.

Accurately predicts mechanical and thermal properties of tobermorite and C-S-H.

Abstract

Tobermorite and Calcium Silicate Hydrate (C-S-H) systems are indispensable cement materials but still lack a satisfactory interatomic potential with both high accuracy and high computational efficiency for better understanding their mechanical performance. Here, we develop a Neuroevolution Machine Learning Potential (NEP) with Ziegler-Biersack-Littmark hybrid framework for tobermorite and C-S-H systems, which conveys unprecedented efficiency in molecular dynamics simulations with substantially reduced training datasets. Our NEP model achieves prediction accuracy comparable to DFT calculations using just around 300 training structures, significantly fewer than other existing machine learning potentials trained for tobermorite. Critically, the GPU-accelerated NEP computations enable scalable simulations of large tobermorite systems, reaching several thousand atoms per GPU card with high efficiency. We demonstrate the NEP's versatility by accurately predicting mechanical properties, phonon density of states, and thermal conductivity of tobermorite. Furthermore, we extend the NEP application to large-scale simulations of amorphous C-S-H, highlighting its potential for comprehensive analysis of structural and mechanical behaviors under various realistic conditions.