| Summary: |
2DoNeuron aims to develop electrical and optically controllable artificial neural networks based on, photonic, low-power consumption and high integration
density 2D-based photonic memristors. This will allow to directly train optical memristors to classify standard handwritten digit images as a first proof-of
principle case study. Our vision is that the assembly of such optical adaptive switches into novel architecture systems will provide a revolutionary paradigm for
heavily distributed, ultrafast data processing with vast impact in neuromorphic computing. By successfully accomplishing the fabrication of an optical neural
network we will give a significant step to pave the way for real-time processing of information encoded by light patterns.
Two hot research topics motivate this highly ambitious goal. The first is the realization of memristors whose dynamical properties mimic synapses [1] and the
behavior of the neuron [2]. The second is the revolution currently taking place in the area of 2D materials, which brings unprecedented physical properties. By
integrating the two, we will pursue a new path toward faster, smaller, and more versatile systems [3].
Memristors show great potential for brain-inspired computing, since they allow in-memory computing [4]. Various materials have been studied as switching
layer, but their overall performance is still a critical issue in the field. To overcome this, one promising approach consists on replacing the typical metallic and/or
insulating layer sandwiched by two electrodes with emerging materials showing enhanced capabilities which, at the same time, could provide new features to
the devices, such as optical tunability and flexibility [5]. Along these lines, 2D materials are at the center of an increasing research due to their unique electronic,
chemical, mechanical and optical properties.
The exploration of 2D materials to emulate synaptic behavior has only recently been attempted [6] and we aim to further  |
Summary
2DoNeuron aims to develop electrical and optically controllable artificial neural networks based on, photonic, low-power consumption and high integration
density 2D-based photonic memristors. This will allow to directly train optical memristors to classify standard handwritten digit images as a first proof-of
principle case study. Our vision is that the assembly of such optical adaptive switches into novel architecture systems will provide a revolutionary paradigm for
heavily distributed, ultrafast data processing with vast impact in neuromorphic computing. By successfully accomplishing the fabrication of an optical neural
network we will give a significant step to pave the way for real-time processing of information encoded by light patterns.
Two hot research topics motivate this highly ambitious goal. The first is the realization of memristors whose dynamical properties mimic synapses [1] and the
behavior of the neuron [2]. The second is the revolution currently taking place in the area of 2D materials, which brings unprecedented physical properties. By
integrating the two, we will pursue a new path toward faster, smaller, and more versatile systems [3].
Memristors show great potential for brain-inspired computing, since they allow in-memory computing [4]. Various materials have been studied as switching
layer, but their overall performance is still a critical issue in the field. To overcome this, one promising approach consists on replacing the typical metallic and/or
insulating layer sandwiched by two electrodes with emerging materials showing enhanced capabilities which, at the same time, could provide new features to
the devices, such as optical tunability and flexibility [5]. Along these lines, 2D materials are at the center of an increasing research due to their unique electronic,
chemical, mechanical and optical properties.
The exploration of 2D materials to emulate synaptic behavior has only recently been attempted [6] and we aim to further explore this path. Memristors will be
developed and studied based on the combination of novel 2D materials. Owing to their atomic-scale thickness, the devices can be scaled down to sub-10 nm
sized, leading to a lower switching voltage (less power consumption) and higher integration density [7].
2D materials such as MoS2, MoSe2, WSe2 and WS2 will allow us to use light as an additional degree of freedom to dynamically tune memristors properties,
adjusting the synaptic learning of individual elements or regions of elements [8]. Using light to read and write the state of a device has the advantage of higher
bandwidth signaling, faster transmission speed, lower crosstalk and decoupling from electronic noise [9].
The production of 2D materials is still a challenge and, besides optimal fabrication processes, other materials are being studied to further improve the
performance. Some of these choices include hybrid stack with polymers [10] creating new interesting interactions and properties. The electrodes also play an
essential role, as they also influence the physics and kinetics of the switching behavior. This is further interesting and challenging when defining dynamical
interfaces with 2D materials and exploring optical influences.
Our work plan will focus on bringing together these innovative findings to achieve the aforementioned 2D-based memristors, switchable by optical and
electrical means. This will take place through a series of connected steps that, at the end, will allow the fabricated systems to be tested under real problems,
paving the way for extraordinary applications in neuromorphic computing, optoelectronics, including optical switches for on-chip photonics, and optical
communication. The partners will start a collaboration already this year for the development of these devices under a Fulbright award. These results will put
Portugal at the forefront of research in novel computing paradigms, leaving it in a privileged position to invest in the related emerging industry opportunities.
This plan requires expertise in materials, nanotechnology, and electronics, and so it will combine the efforts of leading research groups in the fields. IFIMUP has
a large experience on thin film deposition and characterization of nanostructures; UT-Austin provides a strong background on the deposition of 2D materials
down to the ultra-thin (few atomic planes) range, device fabrication and flexible electronics. 2DoNeuron also involves young researchers (PhD and MSc Students)
that will greatly profit from being in direct contact with forefront research and facilities. The main goals of 2DoNeuron can be summarized as:
-To optimize the deposition of stacks of 2D and organic materials
-To characterize the resistive switching and neuromorphic properties of the nanostructures
-To train the devices for the recognition of hand-written letter and digit images |