Study of Field-Coupled Nanocomputing based on molecules for neural systems

Artificial neural networks (ANNs) have made tremendous progress, enabling the achievement of impressive results in artificial intelligence applications. In the last few years, the research highlighted the massive resource requirements of ANNs: contrarily to these, the human brain is capable of performing more general and complex tasks at a minuscule fraction of the power, time, and space required by state-of-the-art supercomputers. In parallel to this in recent years, the scaling process dictated by Moore's Law showed its future limitations, which led to the study and development of new ways to encode information. Among the extended scenario of proposed answers, there is a group of solutions classified as Beyond CMOS technology. In this context, the Field-Coupled Nanocomputing (FCN) is one of the most promising, thanks to its two intrinsic properties: ultra-small devices in large functional-density arrays, and low power dissipation. The problem to be solved is related to the huge power consumption and space required by common ANNs. The research is trying to face the problem, working on the so-called spiking neural network (SNN), which should consume less power and occupy less area, performing the same task. This thesis work joins both ANNs and FCN paradigm, proposing a model for an artificial neuron in the molecular implementation of the FCN paradigm. In this technology, the information is encoded in the charge distribution of a molecule, and the information propagation is enabled by the intermolecular electrostatic interaction, without the need for charge transport. The choice to use the molecular implementation derives from the fact that it has been shown that a molecule has a non-linear behaviour, capable of adding up the effects of surrounding molecules. The fact that a molecule has intrinsically the desired behaviour suggests that simple and compact structures can be obtained at the architectural level. The methodology used to study the topic wants to be as general as possible. So starting from the neuron model, instead of using specific molecules and see if their features fit the goal, it has been chosen to work in terms of molecules transcharacteristics. This allows to define the properties a molecule should have to accomplish its task as a neuron. To simulate the molecular structure, a Self Consistent ElectRostatic Potential Algorithm (SCERPA) is used, this tool analyses in an iterative way the molecule interactions. The first fundamental result obtained in this work is the proof that the proposed structure actually behaves like an artificial neuron. The output switches when the sum of the weighted inputs reaches a certain threshold. In addition to this, the characteristics that the input molecules must have to encode a certain weight have been defined. The properties that the output molecules must have to correctly propagate the information and to define the threshold for the output activation have also been identified. The outcomes obtained confirm what was expected, namely that the molecules, for their intrinsic characteristics, lend themselves well to create an artificial neuron. The basic structure is very simple and compact when compared to the implementation of a "silicon neuron". Moreover, the paradigm used permits to encode and transmit information without the use of current flow, being the information represented as the spatial charge distribution of a molecule, thus reducing the power consumption.