Material characterization of aluminosilicate hydrate geopolymers using deep learning assisted tailor-made potential

Abstract

Aluminosilicate hydrate geopolymers, such as sodium aluminosilicate hydrate and potassium aluminosilicate hydrate, are low-carbon cementitious materials commonly synthesized from industrial solid wastes such as fly ash and slag, offering a sustainable cementitious material with superior mechanical and environmental performance. However, their atomic-scale mechanisms remain elusive due to limitations in experimental and conventional simulation methods. In this study, a machine learning potential model constructed within a deep potential generator framework is developed for sodium aluminosilicate hydrate, and trained on density functional theory datasets spanning from 300 K to 1000 K. The model accurately reproduces density functional theory-calculated energies and forces with errors of 0.005 eV per atom and 0.078 eV Å⁻¹. Furthermore, transfer learning is employed to adapt the sodium aluminosilicate hydrate model to potassium aluminosilicate hydrate using a small amount of additional density functional theory data, yielding comparable accuracy and faster convergence, with errors of 0.003 eV per atom and 0.092 eV Å⁻¹ for energies and forces. To the best of our knowledge, this work presents the first machine-learning interatomic potentials specifically developed for both N-A-S-H and K-A-S-H geopolymer gels. Structural characterization, elastic properties, and dynamic behaviors predicted by the models are benchmarked against density functional theory, classical forcefields, and experimental measurements, demonstrating robustness and transferability of the approach. The findings demonstrate that the method is highly capable of reliably capturing complex aluminosilicate systems, which provide a new atomic-level understanding of their structural and mechanical behavior, thereby establishing a robust basis for guiding the targeted design of durable, high-performance, and sustainable geopolymer materials.