Design of neural networks based on molecular Field-Coupled Nanocomputing

Research related to Artificial Neural Networks (ANN) enhanced the realization of efficient hardware structures able to perform new and highly complex tasks. The hardware realization of ANN requires massive resources, leading to space requirements, computational time, and power consumption challenges. In parallel to this, the scaling process dictated by Moore's law has started to show future limitations related to the same technological aspects. Therefore, researchers began to study and develop new ways of realizing digital electronics. Among the proposed alternatives, molecular Field-Coupled Nanocomputing (FCN) is one of the most promising "Beyond CMOS" technologies. The two main characteristics of this paradigm are the possibility of implementing devices in highly dense functional arrays and the reduced power dissipation. The proposed hardware neural network has its basis in the molecular implementation of the FCN paradigm. In this technology, the information is encoded in the charge distribution within each molecule, whereas electrostatic interactions between neighboring molecules ensure information propagation thus avoiding current flow. Molecular FCN has shown promising results concerning the realization of standard digital gates. The work developed in this thesis has its basis on the results already present in the literature, which show the potential advantages of using FCN to implement single neuron models. These results come from the demonstrated linear behavior of the adopted molecules, which are capable of adding up the effects from the surrounding circuit. Starting from these concepts, this thesis expands the analysis towards realizing new hardware architectural solutions for neural network development. The procedure followed during this work analyses possible strengths and issues of the proposed implementation and consists of two-steps. First, the interface molecules have been characterized to encode specific weights, i.e., coefficients used to increase or decrease an input's relevance, and enlarge their study to cover process variations. Then, the properties of the output molecules and the information propagation structure have also been characterized. Therefore, it was possible to determine the properties the molecules should have to accomplish the task as neurons and derive the layout requirements of the circuit. A Self-Consistent ElectRostatic Potential Algorithm (SCERPA) is adopted to simulate the circuits and solve molecular interactions iteratively. The outcomes confirm the possibility for the proposed structure to work as a neural network. Indeed, each neuron's output switches when the sum of the weighted inputs reaches a defined threshold, and the computed pieces of information propagate correctly through the network. Therefore, it was possible to predict the circuit's final output value with remarkable accuracy. Finally, we demonstrate the functionality of the proposed network to perform pattern recognition tasks and compare the outcomes with those obtained from software-trained feed-forward neural networks. The comparison was successful, and the proposed circuit correctly classified four different 3x3 matrix patterns. Moreover, the molecular FCN network uses fewer neurons to perform the task. This thesis places grounding rules for realizing molecular FCN neural networks; further analysis would be needed to fulfil possible more complex requirements than the ones analysed so far.