RT info:eu-repo/semantics/doctoralThesis
T1 Multiscale modelling of Ge heteroepitaxy on Si
A1 Martín Encinar, Luis
A2 Universidad de Valladolid. Escuela de Doctorado
K1 SiGe
K1 SiGe
K1 SiGe
K1 thin fims
K1 Películas delgadas
K1 Classical Molecular Dynamics
K1 Dinámica Molecular Clásica
K1 Machine Learning
K1 Aprendizaje automático
K1 3312 Tecnología de Materiales
AB The technology of SiGe devices involves the heteroepitaxial growth of SiGe layers with high Ge content or pure Ge layers on Si substrates. The lattice mismatch between Si and Ge generates compressive stress in the grown film that, depending on deposition conditions, mainly releases through the formation of 3D islands or by the nucleation of dislocations. To control the features of the final film, it is important to understand the physics behind the growth process. Some aspects, especially those related to the origin of dislocations and, in general, about the atomic details during the early growth stages, are still unknown.In this Thesis, we investigate the heteroepitaxial growth of Ge on Si from a theoretical perspective, with two main aims: to advance on the current understanding of the processes that occur at atomic level during the initial growth stages, and to develop efficient models able to describe the evolution of the Ge film at the macro-scale. To accomplish these goals, we employ a multiscale simulation framework, which encompasses atomistic techniques such as ab initio and Classical Molecular Dynamics (CMD), together with continuum approaches based on Neural Networks.We first assess the robustness and applicability of CMD for our study, as there is a disparity between experimental and CMD time-scales. Our analysis starts by choosing a suitable potential able to capture the basic properties of Si-Ge systems. After a comparative study, we determine that the Stillinger-Weber potential is the most adequate. We use it in CMD simulations to characterize the relevant atomic mechanisms occurring during early growth stages. In the case of surface diffusion and intermixing, which are interrelated, we had to develop a novel methodology able to disentangle them and concurrently extract their parameters. We also employ ab initio calculations to characterize particular Ge ad-atom configurations and atomic mechanisms. After comparing CMD results with ab initio predictions and experimental observations, we conclude that CMD can be reliably used to study Ge heteroepitaxy on Si. CMD simulation conditions are comparable to that of out-of-equilibrium deposition experiments, where stress mainly releases by dislocation nucleation.Our study extends to simulate with CMD the growth of 40 Ge monolayers on Si under different thermal conditions. We quantify the critical film thickness, intermixing, stress relaxation and surface roughness. At high temperatures, different types of dislocations appear: misfit, threading and Shockley partial. They generate preferentially in valleys among surface islands, in vacancy-rich regions with high atomic disorder and mobility. As different effects occur concurrently during deposition, we designed ad hoc CMD simulations to investigate separately their role on dislocation formation. We carried out a systematic study by introducing regions of different sizes and types (vacancy-rich, amorphous-like, or grooves) on the surface of the Ge film. We find a synergic effect between amorphization and vacancy accumulation that favors plastic relaxation. Dislocations nucleate preferentially at the edges of deep and narrow grooves, and their morphology depends on the groove width.Finally, we explore the use of a Machine Learning approach to simulate the morphological evolution of Ge strained films on Si during growth as an alternative to conventional continuum methods. In strained films, surface diffusion is governed by the elastic energy density contribution, which remains the bottleneck when dealing with time evolutions. We train a convolutional neural network on a dataset comprising film profiles and the corresponding elastic energy density ρ for each profile. The trained model provides quantitative predictions of ρ for arbitrary profiles, surrogating its explicit calculation. The accuracy and robustness of the deep-learned ρ are further demonstrated in the time-integration of surface evolution problems described by partial differential equations of surface diffusion.
YR 2024
FD 2024
LK https://uvadoc.uva.es/handle/10324/76343
UL https://uvadoc.uva.es/handle/10324/76343
LA eng
NO Escuela de Doctorado
DS UVaDOC
RD 19-jul-2025