Investigating the use of Semiconductor Nanowires for Neural Networks

Sammanfattning: In this work, we intend to employ nanowires for the realisation of an artificial neuron which can be used to design a neural network that will guide the new generation of non-von Neumann architectures such as Neuromorphic Computing. Possible applications are strongly related to Neuromorphic architectures and artificial intelligence, such as high-performance computers, robotics hardware and autonomous drones. Here, we aim to design a device that can perform as a node in an optically communicating neural network. We employ nanowires to form a single artificial neuron by coupling a receiver and a transmitter of optical signals, together with a transistor as an in-between control element. Based on this configuration, we find optimal optical properties for our desired neural network communication. The node-to-node communication relies on the strength of the connection known as weights. The connection strength can be tailored by the position and rotation of the individual nanowires making up the nodes and layers; however, this might set some correlations and limitations on the weights. We frame our main research questions as a set of hypotheses that we test by numerical experiments using the finite-difference time-domain (FDTD) method. The FDTD method is one of many numerical methods that exist to tackle this type of electromagnetic simulation. Among the valid alternatives are the finite element method and the method of the moment. However, we decided on the FDTD method because it is the one that best suits our needs to perform electromagnetic simulation on nanowires and to study a node-to-node communication. We split our study and perform separate simulations on the receiver and transmitter. During the simulation of the receiver, we sweep the incoming light angle to study the absorption difference between two distinct depletion regions. The difference in absorption between the two regions yields a potential difference that allows switching on and off the transistor that controls the current through the transmitter. The emitted light from the transmitter is assumed proportional to the current. To study the transmitter, a dipole source was placed in the middle of a nanowire to study and evaluate the ability of the emitter to emit light in a specific direction, namely directivity. The results from the receiver and transmitter studies are combined into a weight function after connecting the two simulated nanowires that form an artificial neuron. The weight function is our ultimate result from our experiment, which can be employed to design a custom neural network based on already known weights.