Design of computing units for the internet of bio-nano things

The Internet of Bio-Nano Things (IoBNT) is a visionary concept for a communication network that aims to connect tiny biocompatible devices called Bio-Nano Things (BNTs) to the Internet. The BNTs can be employed in biological environments, such as the human body, thereby enabling revolutionary opportunities in healthcare. While the connection to the Internet is only established via one interface, the BNTs operate and communicate with each other in an unconventional way, i.e. with molecular signals instead of electrical or electromagnetic signals. As a result, high energy efficiency and biocompatibility are achieved, and operations on the nanoscale are enabled. The main topic of this thesis is nanoscale computing with molecular signals. Besides the overview of existing approaches of computational units, a novel concept capable of emulating a feedforward neural network (FNN) is proposed in this thesis. For this purpose, first a structure of interconnected compartments is introduced, which can be employed to perform a matrix-vector multiplication. The working principle is based on the diffusion of molecules and on chemical reactions occurring in some compartments. If the diffusion process between compartments proceeds much faster than the chemical reactions, the realised matrix coefficients can be set solely via the volumes and contents of the compartments. Otherwise, the coefficients are additionally dependent on the dimensions of the channel connecting the compartments, the diffusion coefficient, and a reaction rate constant. In addition to steady-state investigations, which reveal the proper functioning of the matrix-vector multiplication, the dynamic behaviour is modelled, and estimates for the computation time are derived. Moreover, design guidelines for the structure are provided, and established models are verified using microscopic and mesoscopic simulation methods. Subsequently, the structure is extended to mimic an FNN. Strategies for improving the network design, such as regularisation and pruning, are provided since these techniques are crucial to obtain a practicable compartment-based FNN. To demonstrate the validity of the overall concept, the proposed FNN is employed to solve a regression and a classification task. Finally, interfaces are discussed, which enable the conversion of external into internal molecular signals within the computing unit and vice versa. This conversion can be achieved, for example, by using compartments whose walls contain switchable membranes, allowing molecules to be exchanged upon membrane opening. However, before releasing the analogue result of the computing unit into the external environment with the intention of transmitting it to a desired target, it must be converted to a digital representation due to the higher reliability. To this end, a structure emulating a flash analogue-digital converter is proposed.