Mohr synaptic transistors for homogeneous architecture reserve pool calculations

Research background

Mohr material is a new low-dimensional quantum material system with novel strong correlation and topological physical properties formed by stacking two-dimensional atomic crystals. Unlike traditional quantum materials, Mohr materials have rich quantum states (e.g., associated insulators, orbital magnetism, interfacial ferroelectricity, etc.), which can be regulated by external fields such as electric field, light field, stress field, etc. This makes Mohr materials not only a new ideal platform for physical property exploration, but also shows great potential in the application of Moiré electronic devices. At present, the relevant research mainly focuses on the exploration and regulation of new quantum states in Mohr materials, and how to use the unique quantum states and regulatory laws of Mohr materials to design Mohr electronic devices is a topic that has attracted widespread attention.

Introduction to the content

In the face of the above opportunities and challenges, the collaborative team of Professor Miao Feng of the School of Physics of Nanjing University built a Moiré synaptic transistor that simulates the short-term plasticity and long-term plasticity of biological synapses in the way of “atomic Lego”, realizing for the first time a Moiré synaptic transistor that can simulate the short-term plasticity and long-term plasticity of biological synapses. Further, based on the highly tunable characteristics of the dynamics of the Moiré synaptic transistor, the collaborative team proposed a full-moiré physical neural network (MPNN) that can perform homogeneous architecture reserve pool calculations.

Figure 1. Moiré synaptic transistor. (a) Schematic diagram of biological neural networks and synapses. The intensity of the spike signal sent by neurons affects the changing behavior of postsynaptic membrane current, with weak signals triggering short-term plastic behavior (STP) and strong signals stimulating long-term plastic behavior (LTP). (b) Schematic diagram of a molar potential field with adjustable electric field generated by the alignment of hexagonal boron nitride with the graphene lattice. Under the action of weak electric field, the shallow Moiré potential realizes short-time modulation of the electron distribution. Under the action of strong electric field, the deep Moiré potential strongly localizes the electrons to the upper layer of graphene to achieve long-term modulation. (c-d) Schematic diagram and optical diagram of a Moiré heterojunction-based synaptic transistor. (e) Transfer characteristic curve of Moiré synaptic transistors over the ±4 V scan range. (f) The transfer characteristic curve of the Moiré synaptic transistor at the ±12 V scanning range, showing a distinct memory window. (g) Memory window size vs. sweep voltage range.

Figure 2. Experimental characterization of short-term plasticity. (a) Simulation of short-term memory behavior by Moiré synaptic transistors. (b-c) Simulation of double-pulse alienation behavior by Moiré synaptic transistors and the variation of double-pulse alienation index with applied pulse interval. (d-f) Under external stimuli of different intensities (changing the amplitude, width, and number of applied pulses), the dynamic current changes of the Moiré synaptic transistors exhibit distinguishable dynamic behavior and can be used to construct physical reserve pools.

Figure 3. Experimental characterization of long-term plasticity. (a-b) Simulation of long-term enhancement and long-term inhibition behavior by Moiré synaptic transistors, exhibiting long-term memory properties. (c) Moiré synaptic transistors are continuously regulated based on conductance under the action of pulse trains. (d) Retention time characterization of the 8 conductance states selected by the Moiré synaptic transistor. Each state of the device does not degrade significantly for 1000 seconds, which can be used as a typical neural network weight device.

Figure 4. Full Moiré Physical Neural Network. (a-b) Schematic diagram of a Moiré physical neural network for a reserve pool computing system performing a homogeneous architecture, comprising a reserve pool layer and a readout layer constructed from Moiré synaptic devices of the same class. (b) Taking the encoded “1100” and “0011” time domain information pulse sequences as an example, the processing of time domain information by Moiré synaptic devices as reserve pool nodes is demonstrated, where the current values read at different sampling moments can be used as reserve pool states. (c) Distribution of pool status under 16 input sequences read out at two different sampling moments. (d) Weight state distribution after readout layer training is completed. (e) The relationship between the training period and the recognition rate of the reserve pool state obtained based on different sampling modes. The recognition rate can reach 90.8% based on mixed sampling mode. (f) The recognition results of the test set after the training is completed.

Research significance and importance

This work is the first attempt of graphene Mohr materials for neuromorphic calculations, which opens up an important technical route for the development of Mohr electronics. (Source: ChinesePhysicsLetters WeChat public account)

Related paper information:

Special statement: This article is reproduced only for the need to disseminate information, and does not mean to represent the views of this website or confirm the authenticity of its content; If other media, websites or individuals reprint and use from this website, they must retain the “source” indicated on this website and bear their own legal responsibilities such as copyright; If the author does not wish to be reprinted or contact the reprint fee, please contact us.

Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button