Researchers develop a convolutional neural network architecture to predict the electrical properties of the material when strain is applied.
Nanoscale semiconductor devices are reaching their limit of performance. To get further performance, it is necessary to modify the electrical properties of the material for high electron mobility. Elastic Strain engineering (ESE) is one such field that deals with straining materials to improve their performance. Adding mechanical stress or strain to the material can change its electrical properties. Like applying stress to silicon material can result in smaller transistors with high electron mobility and also enable more efficient solar cells.
Materials at nanoscale can withstand major deformation. Materials in strained state can exhibit remarkable optical, thermal, electronic, and other properties due to a change in interatomic distances. Moreover, the properties of the materials can be further modified by varying the strain level. For these modifications, Skoltech researchers and their colleagues from the U.S. and Singapore have created a neural network that can help modify semiconductor crystals in a controlled way to achieve superior properties for electronics. The paper was published in the journal npj Computational Materials.
The team of researchers earlier used ESE to turn nanoscale diamond needles from insulating to highly conductive and metal-like. This development provides an idea how impactful this technology can be. Now the team of researchers have created a neural network architecture that can aid designers and scientists in their ESE efforts for semiconductors.
The convolutional neural network designed here takes strain tensor as an input and predicts the electronic band structure of the material accordingly. This provides insight on the material’s electrical property and it can be used to calculate any properties of interest, including the bandgap, its properties, and electron effective mass tensor. To improve this model, the researchers have also taken into account the physical properties and symmetries of the material.
The model combines various data sources and has an active learning feature. Active learning allows the model to guess what data may be the most useful to obtain in the next training stage, and use it for training. In the final stage, the network is trained on a set of computationally expensive data from the very accurate GW-based calculations. This learning procedure allows to reduce the amount of computations needed
The team claims that its new neural network is “more versatile, accurate, and efficient in its capacity to facilitate autonomous deep learning of the electronic band structure of crystalline solids” than state-of-the-art solutions.
The researchers say about their upcoming studies, “We are currently working on our next paper, which is devoted to the boundaries of admissible elastic strains. It is an important topic since the theoretical limits of safe elastic deformation for ESE are yet to be discovered.”