Redes neurais artificiais para predição das propriedades de transporte de vale em nanofitas de grafeno

• Autor(a) principal: Silva, Pablo Rhuam Cavalcante
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do Mackenzie
• Texto Completo: http://dspace.mackenzie.br/handle/10899/24502
• topic: rede neural; aprendizado de Máquina; valleytronics; deformação mecânica; bolhas gaussianas; filtro de vale; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO::COMPUTABILIDADE E MODELOS DE COMPUTACAO

The control of the valley degree of freedom of electrons (valleytronics) has recently emerged as a promising technology for the next generation of electronic devices; this quantum number naturally appears in periodic solids with degenerated local minima and maxima at inequivalent points of the Brillouin zone. Similar to spintronics, the applicability of valleytronics relies on the electric generation, control and detection of valley currents; in this context, 2D materials with hexagonal lattices such as graphene and transition metal dichalcogenides offer two valleys (K and K0) well separated in momentum space that can be accessed by optical, magnetic and mechanical means. Electrons in opposite valleys in graphene see in homogeneous mechanical deformation as regions with opposite polarity pseudo-magnetic fields; unfortunately, the observed effects require fine tuning of the energy, defined height/width ratio of the bubble, narrow contacts, location of the nanobubble near to the right contact and crystalline orientation. All these ingredients mix in a subtle and hidden way that it is impossible to predict the effect of a given deformation before heavy calculations. With this in mind, this problem was modeled as a supervised regression problem in which a Deep Neural Network (DNN) model was trained to learn the relationship between the design parameters of the superlattice and the valley filter effect. The dataset used in this experiment was generated using the Green’s Function. The dataset was processed, normalized and randomly split into 80% for training and 20% for testing (used to evaluate the performance of the DNN model). The results using DNN for the prediction showed accuracy measures in: R2 = 0:970 and MSE = 0:001. The DNN was able to predict the valley filter with a precision similar to the Green’s function but with much less computational effort. The DNN was then used to help in the searching for the 1D Gaussian superlattice with efficient valley filter, but with the smallest number of deformations and strain.