Morphological evolution of films and nanostructures by state-of-the-art Machine Learning approaches

Atomistic simulations of growth at typical experimental conditions face some critical challenges due to both the extremely long time-scales and the large (often even when dealing with nanostructures!) size scales involved. Continuum approaches, while yielding only an average description of the system behavior, can help closing the aforementioned scale gap, as we shall show by discussing several examples [see, e.g., Ref. 1]. However, especially when dealing with anisotropic systems and/or in the presence of stress fields, continuum models can also be computationally demanding. The fast development of Machine Learning (ML) approaches in the last few years influenced all scientific fields and did not spare modeling of phenomena relevant to crystal growth. If on one hand attention was largely focused on the development of a new generation of accurate but relatively cheap interatomic potentials, techniques such as convolutional, recurrent neural networks have been exploited by us and other groups to speed up continuum simulations. We shall show, by mainly focusing on surface-diffusion-driven deterministic evolution [2], that reliable predictions are possible on extended time scales and that uncertainty can be conveniently quantified. Interestingly, ML models are not limited to deterministic dynamics. Very recently, indeed, we showed that generative adversarial networks can be exploited to learn stochastic processes [3]. Overall, it is evident that a new generation of ML-based methods for investigating growth is presently under development, opening exciting perspectives in terms of process control and efficient choice of the growth parameters.