凯发

Artificial Intelligence-based techniques have become critical to accelerate the innovation processes from material growth to characterization and properties prediction. Particularly challenging is the modelling of disordered (amorphous) materials, complex interfaces and random assemblies of (2D) materials which are mainly used today in applications.
In that perspective, amorphous materials such as boron nitride (aBN) and “amorphous graphene” (aG) have recently become prominent materials for many different applications due to their good properties such as thermal stability, mechanical properties, insulating behaviour, and ultralow dielectric constant in aBN (<2). Moreover, amorphous films are more suitable to large area deposition compared to clean hBN or graphene since they can be grown at low temperatures (about 400 ºC) and on various substrates [1-3]. However, their properties depend on the nature and degree of disorder, which needs a well-defined metrics for benchmarking different materials. Having such metrics in place will allow to tune the properties and performance of these films during the fabrication for desired applications. In this context, revealing the relationship between fabrication strategies and the material properties of the film is also crucial.

Capturing the key features of the amorphous nature of materials requires theoretical characterization to understand how material properties change with the microstructure. Since simulations of amorphous materials need large structural models, density functional theory (DFT) is not a suitable tool despite the high accuracy it offers. On the other hand, molecular dynamics (MD) simulations with empirical interatomic potentials require much less computational cost; however, they can turn out to be not accurate enough to correctly describe the local environment of amorphous materials. Machine learning-driven interatomic potentials (ML-IP) can describe the local environment with a similar accuracy to DFT and at a much lower cost [4,5]. Here, we introduce Gaussian approximation potentials (GAP) for atomistic simulations of aBN incorporating different contaminators and doping materials, which are trained on a large dataset of atomic structures generated by DFT calculations [6-8]. We will present a systematic analysis to screen out possible realistic morphologies as a function of growth parameters, such as temperature, quenching rate, and the presence of a dopant, and their corresponding material properties using GAP-driven MD simulations. The extensive simulations of a large quantity of possible structures presented here can guide experimental research and provide trends of scaling behaviour as a function of experimentally controllable parameters. The impact of amorphousness on dielectric properties will be also discussed for aBN and aG in the light of recent breakthroughs and claims [9,10].

References
[1] Hong, S et al. Nature 582, 511–514 (2020).
[2] Glavin, N. R., et al., Adv. Func. Mat. (2016).
[3] Chen, C. Y., et al., arxiv: 2312.09136 (cond-mat).
[4] Unruh D. et al. Physical Review Materials 6, 065603 (2022).
[5] Deringer, VL. and Csányi, G. Physical Review B 95, 094203 (2017).
[6] Kaya, O., et al. Nanoscale Horizons 8, 361–367 (2023).
[7] Kaya, O. et al. J. Phys. Materials 7, 025010 (2024).
[8] Kaya, O. et al. arXiv:2402.01251 (cond-mat).
[9] Th. Galvani et al. Journal of Physics: Materials 7 (3), 035003 (2024); arXiv:2403.11924 (cond-mat).
[10] Th. Galvani, O. Kaya, S. Roche, unpublished