Artificial neural networks have revolutionized the field of artificial intelligence with human-like performance in fields such as computer vision and speech recognition already achieved. The ANNs take inspiration from the operation of biological neural networks, e.g. the human brain by employing layers of artificial neurons connected to each other via synaptic elements characterized by a weight factor. Yet, the von Neumann architecture of the present computing systems use opposite strategies compared to the brain: centralized versus distributed memory, and always-on versus event-driven systems. In practice this results in the software driven ANNs being orders of magnitude less energy efficient than the human brain, and the continuously growing demand for new machine learning models of increasing complexity account for a substantial part of the global energy consumption.
Neuromorphic computing presents a potential solution to these challenges by emulating the functionality and interconnectivity of the biological neural networks directly on hardware level. However, purely CMOS based neuromorphic circuits are impractical for the implementation of large networks, as even a single synapse or neuron can take tens of transistors each to implement. To accelerate the development of neuromorphic circuits, a next-generation device technology is needed. In practice, this equates to emulating the key features of the neurons and synapses directly on the materials’ level.
For the synaptic connections, a potential new technology are memristors that can emulate the synaptic weight through their history-dependent variable conductance. Various inorganic materials have been employed for these resistive switching devices that rely on the formation/dissolution of a conductive filament within an insulating matrix. The main challenges preventing widescale implementation are the device-to-device and cycle-to-cycle variability arising from the stochastic nature of the filament formation. Additionally, the typically high conductance in the ON-state results in high energy consumption.
The aim of the RESWITCH project is to enable a new type of resistive switching device concept based on metal-organic coordination polymer thin films. With redox-active ligands, the emulation of the synaptic weight can be based on the oxidation-state dependent properties of these thin films as the electronic conductivity can be modulated electrochemically with dynamic operation arising from the concurrent counter-ion motion. The interplay of the metal and organic constituents allows for precise control of the electric/electrochemical properties, but poor processability has been limiting their applicability for nanotechnology applications. The key element in RESWITCH is to implement a new thin film -based approach for conductive coordination polymers using Molecular Layer Deposition (MLD), a vapor phase thin film deposition method derived from the Atomic Layer Deposition (ALD) technique. As with ALD, it is defined by the sequential, self-saturating exposure of vapor-phase precursors onto surfaces. This monolayer accuracy in process control leads to sub-nanometer range precision in layer thickness control and in excellent uniformity over large area substrates,
The Objectives of the Project are three-fold:
1. Establishing novel MLD-chemistries based on redox-active organic ligands.
2. Gaining detailed understanding on the link between the thin film composition and it’s redox-properties and electrical conductivity.
3. Integration of the newly developed materials in a resistive switching device demonstrator with artificial synapse -like functionality.
