Skoltech researchers and their partners in the U.S. have created a neural network that can help tweak semiconductor crystals to achieve superior properties for electronics. This is an exciting new direction of development with limitless possibilities for next-generation chips and solar cells. This study is published as a paper in the journal npj Computational Materials.
Nanoscale materials can sustain significant deformation. Due to a change in interatomic distances, they can exhibit amazing optical, thermal, electrical, and other capabilities in what’s known as the strained state. A stretched material’s intrinsic properties can change, such as semiconducting silicon converting into a material that conducts electric current freely.
Additionally, these properties can be changed on demand by adjusting the strain level. Elastic strain engineering, or ESE, is a new branch of study based on that concept. When all other alternatives for improving chip performance have been exhausted, the approach could be used to modify semiconductor performance, perhaps enabling a workaround for the impending Moore’s law limit. Another application can lie within the field of solar cell development.
Previously the researchers used ESE to turn nanoscale diamond needles from insulating to highly conductive and metal-like, showing the range of possibilities with this technology. The team has now developed a convolutional neural network architecture that may be used to direct ESE efforts in semiconductors.
About the study
The strain tensor is fed into the neural network, which predicts the electronic band structure—a physical ‘snapshot’ that classifies the electronic properties of stressed material. It can then be used to compute any desired properties, such as the bandgap, its properties, and the electron-effective mass tensor.
The method combines multiple data sources, such as computationally cheap but inaccurate data with precise but expensive data, to improve the model’s accuracy and convergence. Another unique aspect is active learning, in which the model is allowed to predict which data will be most beneficial to gather in the next training stage and then use that data for training.
The network is then trained using a set of computationally expensive data derived from the highly accurate GW-based calculations, which helps to reduce the number of computations required.
According to the researchers, the new neural network is “more adaptable, accurate, and efficient in its potential to permit autonomous deep learning of the electrical band structure of crystalline solids” than existing systems. This makes searching and optimizing inside the strain space faster and more accurate, resulting in the best strain values for given figures of merit.
This study is actually part of a year-long collaboration between Skoltech, MIT, and Nanyang Technological University. The Skoltech scientists focused on the computational and machine learning aspect and their colleagues handling the physical aspect of the work. They will now be working on the boundaries of admissible elastic strains. It’s a crucial topic because the theoretical limits of safe elastic deformation for ESE are yet unknown and to be discovered.