**5. Photonic chip with photon synapse**

**4.1. 8-Level memristor system with an MoS2**

80 Advances in Memristor Neural Networks – Modeling and Applications

the dark or in white light (**Figure 17(a)**).

**Figure 17.** The operation of the MoS2

LRSD6, HRSL6 and LRSL6, which can be read in the dark or in the light [20].

tion of the MoS2

The diagram of the operation of the 8-level memristor system with the MoS<sup>2</sup>

detector is shown in **Figure 17**, where the resistance states formed after the SET/RESET opera-

light are shown. A memristor polarized at 3 V in darkness or in white light demonstrates four states that are read at a voltage of 0.7 V (HRSD3 and LRSD3) and 1.2 V (HRSL3 and LRSL3) in

Polarization of the memristor at 6 V in darkness or under light leads to the formation of four more states that are read at a voltage of 0.7 V (HRSD6 and LRSD6) and 4 V (HRSL6 and

 **floating photogate**

memristor polarized at voltages of 3 V and 6 V in the dark or when excited by

photomemristor, polarized at different voltages in the dark or when excited by

light. (1) high and low resistive states obtained using SET/RESET operations at −3 V/3 V and −6 V/+6 V in the dark (HRSD3, LRSD3 and HRSD6, LRSD6) and under white light (HRSL3, LRSL3, LRSL6 and HRSL6). (2) reading diagram under impulse voltage. Resistive states are read at 0.7 V (HRSD3, LRSD3, HRSD6 and LRSD6), 1.2 V (HRSL3 and LRSL3) and 4 V (LRSL6 and HRSL6) in dark (D) or white light (L). (3) excitation scheme by pulses of white light. SET/RESET and the READ operation is controlled by switching off the light pulses (black) (HRSD3, LRSD3, HRSD6 and LRSD6) and turned on (blue) (HRSL3, LRSL3, HRSL6 and LRSL6). A 3 V polarized memristor demonstrates four states that are read as HRSD3, LRSD3, HRSL3 and LRSL3, while a memristor polarized at 6 V demonstrates the other four states: HRSD6,

floating photo-

A photonic chip containing 70 photon synapses was demonstrated in 2017 by a team from the universities of Oxford, Münster and Exeter [21]. The recording, erasure and reading of information in this case are carried out completely by optical methods (**Figure 18**). The photon synapse consists of a cone-shaped waveguide (dark blue) with discrete islands of phase-change

**Figure 18.** Photon synapse on a crystal. (A) The structure of the neuron and the synapse. Insert: Illustration of the synapse junction. (B) Scheme of the integrated photon synapse resembling the function of a neural synapse. The synapse is based on a cone-shaped waveguide (dark blue) with discrete PCM islands from the top, optically connecting presynaptic (preneural) and postsynaptic (postneural) signals. The red open circle is a circulator with port 2 and port 3, connecting the synapse and postneuron; weighing pulses are fed through port 1 to the synapse. (C) An optical microscope image of a device with an active region (red rectangle) as a photon synapse. The optical input and output of the device is carried out through apodized diffraction couplers (white rectangles). Box: A typical photonic chip containing 70 photon synapses is smaller than a coin. (D) Scanning electron microscope image of the photon synapse active region corresponding to the red rectangle in (C) with six Ge2 Sb2 Te5 (GST) strips (1 × 3 μm, yellow) at the tip of the waveguide (blue). Insert: An increased conical waveguide structure, marked with a white dotted frame [21].

material (PCM) from the top optically connecting the presynaptic (preneuronal) and postsynaptic (postneuronal) signals. The use of purely optical means provides ultrafast operation speed, virtually unlimited bandwidth and no loss of electrical power on interconnects. It is significant that the synaptic weight can be randomly installed simply by changing the number of optical pulses that create a system with continuously changing synaptic plasticity, reflecting the true analog nature of the biological synapses.

**6. 2D Transition metal dichalcogenides (MoS2**

**6.1. Atomistor: nonvolatile atomic resistive TMD memory**

, MoSe2

, WS2

MOCVD growth of semiconductor monolayer MoS2

structure for 2D memristors (**Figure 20**).

**WSe2**

**) memory**

resistive 2D TMD (MoS2

darker areas are covered with MoS<sup>2</sup>

(i) on SiO2

energies of 1.9 eV (h) and 2.0 eV (i). Scale mark, 10 microns [22].

and WS2

and WS2

and WS2

(h) and WS2

MoS2

MoS2

MoS2

**, MoSe2**

silicon oxide at 500°C on a 4-inch wafer allows to obtain excellent electrical characteristics and

In 2017, the Argonne National Laboratory demonstrated an atomistor: a nonvolatile atomic

nanometer [23]. New device concepts in nonvolatile flexible memory and brain-like (neuromorphic) computing can significantly benefit from the tremendous possibilities for designing

and WSe2

**Figure 20.** Single-layer transition metal dichalcogenides (TMD) films on 4-inch wafers. a, b, photos of MoS<sup>2</sup>

show a quartz substrate for comparison. (c) Photo of a patterned monolayer MoS<sup>2</sup>

(b) monolayers of films grown on 4-inch substrates with diagrams of their respective atomic structures. The left halves

samples (diamonds). (g) SEM image and photoluminescence (PL) (bottom insert, at 1.9 eV) of monolayer (ML) MoS<sup>2</sup> membranes suspended on a SiN TEM mesh with holes of 2 μm (the suspended film scheme is shown in the upper inset). Label, 10 microns. (h), (i) Optical images (normalized to the area of a clean substrate) of the patterned monolayer

(orange line) in the photon energy range from 1.6 to 2.7 eV. (e) The Raman spectra of the grown monolayer

normalized to the intensity of the silicon peak. (f) Normalized photoluminescence spectra of monolayers

grown. The peak positions in d–f are consistent with the positions of the peaks obtained from the peeled

). (d) Optical absorption spectra of the MOCVD-grown monolayer MoS2

taken from films with a wafer-scale pattern. The insets show photoluminescent images for

**, WS2**

Memristive Systems Based on Two-Dimensional Materials

films and tungsten disulfide (WS<sup>2</sup>

) memory (**Figure 21**), which scales to a sub-

 **and** 

http://dx.doi.org/10.5772/intechopen.78973

) on

83

(a) and WS2

/Si wafer (the

(red line)

film on a 4-inch SiO<sup>2</sup>

#### **5.1. Synaptic weight and plasticity**

Synaptic adjustment of the device when switching between crystalline and amorphous states of GST islands with a recorded change in the relative transmission coefficient is shown in **Figure 19**. Five weight states of the photon synapse are obtained by switching the energy of the optical pulse (404.5 pJ, 50 ns). The photon synapse demonstrates good reproducibility of weight numbers with cyclic measurements (**Figure 19(B)**). In this case, the photon synaptic weight is determined by the number of optical pulses (**Figure 19(C)**).

**Figure 19.** Synaptic weight and plasticity. (A) Demonstration of the differential synaptic weight of the device in **Figure 18** when switching between crystalline and amorphous GST island states with recorded relative coefficient change (ΔT/ To ). Each weight can be achieved with the same number of pulses (50 ns at 243 pJ, 1 MHz) from any previous weight. (B) Weight repeatability for several cycles. Box: Statistical analysis of the change in readings for the weight "0," "1" and "4". The applied pulse was 50 ps at 320 pJ, slightly larger than in (A). (C) Five weights of the photon synapse are obtained when the energy of the optical pulse is switched (404.5 pJ, 50 ns). Dotted blue (yellow) rectangles correspond to the first (last) weight cycle. The up and down arrows in the rectangles are the weighing directions. (D) Photon synaptic weight (ΔT/T<sup>o</sup> ) as a function of the number of optical pulses. The left (right) panel corresponds to the data of the marked blue (yellow) field in (C). Painted triangles (not filled squares) represent data from the upward (downward) direction of weighing. The dashed lines represent the exponential curves closest to the experimental data [21].

#### **6. 2D Transition metal dichalcogenides (MoS2 , MoSe2 , WS2 and WSe2 ) memory**

MOCVD growth of semiconductor monolayer MoS2 films and tungsten disulfide (WS<sup>2</sup> ) on silicon oxide at 500°C on a 4-inch wafer allows to obtain excellent electrical characteristics and structure for 2D memristors (**Figure 20**).

#### **6.1. Atomistor: nonvolatile atomic resistive TMD memory**

material (PCM) from the top optically connecting the presynaptic (preneuronal) and postsynaptic (postneuronal) signals. The use of purely optical means provides ultrafast operation speed, virtually unlimited bandwidth and no loss of electrical power on interconnects. It is significant that the synaptic weight can be randomly installed simply by changing the number of optical pulses that create a system with continuously changing synaptic plasticity, reflect-

Synaptic adjustment of the device when switching between crystalline and amorphous states of GST islands with a recorded change in the relative transmission coefficient is shown in **Figure 19**. Five weight states of the photon synapse are obtained by switching the energy of the optical pulse (404.5 pJ, 50 ns). The photon synapse demonstrates good reproducibility of weight numbers with cyclic measurements (**Figure 19(B)**). In this case, the photon synaptic

**Figure 19.** Synaptic weight and plasticity. (A) Demonstration of the differential synaptic weight of the device in **Figure 18** when switching between crystalline and amorphous GST island states with recorded relative coefficient change (ΔT/

). Each weight can be achieved with the same number of pulses (50 ns at 243 pJ, 1 MHz) from any previous weight. (B) Weight repeatability for several cycles. Box: Statistical analysis of the change in readings for the weight "0," "1" and "4". The applied pulse was 50 ps at 320 pJ, slightly larger than in (A). (C) Five weights of the photon synapse are obtained when the energy of the optical pulse is switched (404.5 pJ, 50 ns). Dotted blue (yellow) rectangles correspond to the first (last) weight cycle. The up and down arrows in the rectangles are the weighing directions. (D) Photon synaptic

blue (yellow) field in (C). Painted triangles (not filled squares) represent data from the upward (downward) direction of

weighing. The dashed lines represent the exponential curves closest to the experimental data [21].

) as a function of the number of optical pulses. The left (right) panel corresponds to the data of the marked

ing the true analog nature of the biological synapses.

82 Advances in Memristor Neural Networks – Modeling and Applications

weight is determined by the number of optical pulses (**Figure 19(C)**).

**5.1. Synaptic weight and plasticity**

To

weight (ΔT/T<sup>o</sup>

In 2017, the Argonne National Laboratory demonstrated an atomistor: a nonvolatile atomic resistive 2D TMD (MoS2 , MoSe2 , WS2 and WSe2 ) memory (**Figure 21**), which scales to a subnanometer [23]. New device concepts in nonvolatile flexible memory and brain-like (neuromorphic) computing can significantly benefit from the tremendous possibilities for designing

**Figure 20.** Single-layer transition metal dichalcogenides (TMD) films on 4-inch wafers. a, b, photos of MoS<sup>2</sup> (a) and WS2 (b) monolayers of films grown on 4-inch substrates with diagrams of their respective atomic structures. The left halves show a quartz substrate for comparison. (c) Photo of a patterned monolayer MoS<sup>2</sup> film on a 4-inch SiO<sup>2</sup> /Si wafer (the darker areas are covered with MoS<sup>2</sup> ). (d) Optical absorption spectra of the MOCVD-grown monolayer MoS2 (red line) and WS2 (orange line) in the photon energy range from 1.6 to 2.7 eV. (e) The Raman spectra of the grown monolayer MoS2 and WS2 normalized to the intensity of the silicon peak. (f) Normalized photoluminescence spectra of monolayers MoS2 and WS2 grown. The peak positions in d–f are consistent with the positions of the peaks obtained from the peeled samples (diamonds). (g) SEM image and photoluminescence (PL) (bottom insert, at 1.9 eV) of monolayer (ML) MoS<sup>2</sup> membranes suspended on a SiN TEM mesh with holes of 2 μm (the suspended film scheme is shown in the upper inset). Label, 10 microns. (h), (i) Optical images (normalized to the area of a clean substrate) of the patterned monolayer MoS2 (h) and WS2 (i) on SiO2 taken from films with a wafer-scale pattern. The insets show photoluminescent images for energies of 1.9 eV (h) and 2.0 eV (i). Scale mark, 10 microns [22].

requirements for high-temperature phase melting and long switching times have limited their use. 2D memristors offer unprecedented advancement for high-frequency systems due to their low voltage operation, small form-factor, high switching speed and low temperature integration compatible with Si or flexible substrates. Nonvolatile RF switches show promising results with acceptable insertion loss of ~ 1 dB and isolation of >12 dB up to 50 GHz (**Figure 22(d)**). The extracted resistance when the state is On, RON ≈ 11 ohms and capacitance when the state is Off, COFF ≈ 7.7 fF. This results in a cut-off frequency, which is used to estimate the RF switches (a figure of merit (FOM)) [29, 30] fco = 1/(2πRONCOFF) ≈ 1.8 THz. Further improvements, especially in terms of scaling, are expected to lead to a significant increase in FOM. A unique combination of independent LRS resistance and area-dependent HRS capacity gives a FOM that can be scaled to 100 s of THz by reducing the area of the device that determines advantages over phase-change switches [29, 30], where the capacitance is proportional to the width, but RON is inversely dependent, hence, prevents frequency scaling without significant compromise losses. In addition, the high stress of mechanical rupture and the easy integration of 2D materials onto soft substrates enable the production of flexible nonvolatile digital and analog/RF switches capable of withstanding mechanical cycling

Memristive Systems Based on Two-Dimensional Materials

http://dx.doi.org/10.5772/intechopen.78973

85

(**Figure 22(e)**).

**Figure 22.** Characteristics of the atomistor. (a, b) Resistance spread of MoS2

at room temperature. Resistance of HRS and LRS is determined by measuring the current at a small bias voltage of

50 GHz. The cut-off frequency is ~1.8 THz. (e) Stable resistance of states with high resistance and low resistance after

monolayer show promising characteristics with an insertion loss of ~1 dB and isolation >12 dB up to

is 2 × 2 μm<sup>2</sup>

switching cycles. (c) Time-dependent measurements of the MoS2

0.1 V. The area of this transverse device 2 L-MoS2

1000 cycles of bending at 1% strain [23].

1 × 1 μm<sup>2</sup>

MoS2

crossbar MIM devices for 150 manual dc

switch with stable storage of information for a week

. (d) Experimental, nonvolatile RF switches based on a

**Figure 21.** Scheme of a TMD sandwich based on MoS2 grown on Au foil (left) and a representative curve of I-V behavior of bipolar resistive switching in a MoS2 monolayer with a lateral area of 2 × 2 μm<sup>2</sup> (on the right). Step 1: The voltage increases from 0 to 1.2 V. At ~ 1 V, the current rises sharply to the limiting current, indicating the transition (SET) from the high resistance state (HRS) to the low resistance state (LRS). Step 2: The voltage decreases from 1.2 to 0 V. The device remains in the LRS. Step 3: The voltage increases from 0 to 1.5 V. At −1.25 V, the current drastically decreases, indicating a transition (RESET) from LRS to HRS. Step 4: The voltage decreases from −1.5 to 0 V. The device remains in HRS mode until the next cycle [23].

2D materials. A new large application, a static radio frequency (RF) switching, was demonstrated using a MoS2 monolayer operating at 50 GHz.

Multilayer atomic materials [24] can be used to construct the elemental base of neuromorphic computers. One of the new directions is the creation of solid-state memory of the next generation with phase changes and TMO devices. The devices from 2D crystals have certain advantages in obtaining vertical scaling up to the atomic layer. When replacing metal electrodes with graphene, the entire memory cell can be scaled below 2 nm. In addition, the transparency of graphene and the unique spectroscopic features of 2D materials make it possible to obtain a direct optical characteristic of the device on the production line. At present, manual testing of the device's durability (**Figure 22(a)** and (**b)**) is not enough to meet the requirements for solid-state memory and is a reflection of the emerging state of 2D atomistors in comparison with TMO memories [25]. Through engineering or doping, the durability of the device can be improved, similar to what was observed for amorphous carbon storage devices [26]. Retention of nonvolatile states tested up to a week (**Figure 22(c)**) is already sufficient for certain neuromorphic applications with short-term and medium-term plasticity [27]. The subnanometric thickness of monolayers is promising for the realization of ultrahigh densities. With a free step of 10 nm, the atomic density of 1015/mm3 would provide sufficient space to simulate the density of human synapses (~10<sup>9</sup> /mm3 ) [28]. For a single-bit single-level storage device, this corresponds to a theoretical surface density of 6.4 Tbit/inch<sup>2</sup> .

#### **6.2. High-frequency 2D MoS2 memristors**

Modern switches are implemented using transistor or microelectromechanical devices, both of which are volatile, and the latter also requires a large switching voltage that is not suitable for mobile technologies. Recently, phase change switches have attracted interest [29], but the requirements for high-temperature phase melting and long switching times have limited their use. 2D memristors offer unprecedented advancement for high-frequency systems due to their low voltage operation, small form-factor, high switching speed and low temperature integration compatible with Si or flexible substrates. Nonvolatile RF switches show promising results with acceptable insertion loss of ~ 1 dB and isolation of >12 dB up to 50 GHz (**Figure 22(d)**). The extracted resistance when the state is On, RON ≈ 11 ohms and capacitance when the state is Off, COFF ≈ 7.7 fF. This results in a cut-off frequency, which is used to estimate the RF switches (a figure of merit (FOM)) [29, 30] fco = 1/(2πRONCOFF) ≈ 1.8 THz. Further improvements, especially in terms of scaling, are expected to lead to a significant increase in FOM. A unique combination of independent LRS resistance and area-dependent HRS capacity gives a FOM that can be scaled to 100 s of THz by reducing the area of the device that determines advantages over phase-change switches [29, 30], where the capacitance is proportional to the width, but RON is inversely dependent, hence, prevents frequency scaling without significant compromise losses. In addition, the high stress of mechanical rupture and the easy integration of 2D materials onto soft substrates enable the production of flexible nonvolatile digital and analog/RF switches capable of withstanding mechanical cycling (**Figure 22(e)**).

2D materials. A new large application, a static radio frequency (RF) switching, was demon-

monolayer with a lateral area of 2 × 2 μm<sup>2</sup>

increases from 0 to 1.2 V. At ~ 1 V, the current rises sharply to the limiting current, indicating the transition (SET) from the high resistance state (HRS) to the low resistance state (LRS). Step 2: The voltage decreases from 1.2 to 0 V. The device remains in the LRS. Step 3: The voltage increases from 0 to 1.5 V. At −1.25 V, the current drastically decreases, indicating a transition (RESET) from LRS to HRS. Step 4: The voltage decreases from −1.5 to 0 V. The device remains in HRS mode

grown on Au foil (left) and a representative curve of I-V behavior

(on the right). Step 1: The voltage

Multilayer atomic materials [24] can be used to construct the elemental base of neuromorphic computers. One of the new directions is the creation of solid-state memory of the next generation with phase changes and TMO devices. The devices from 2D crystals have certain advantages in obtaining vertical scaling up to the atomic layer. When replacing metal electrodes with graphene, the entire memory cell can be scaled below 2 nm. In addition, the transparency of graphene and the unique spectroscopic features of 2D materials make it possible to obtain a direct optical characteristic of the device on the production line. At present, manual testing of the device's durability (**Figure 22(a)** and (**b)**) is not enough to meet the requirements for solid-state memory and is a reflection of the emerging state of 2D atomistors in comparison with TMO memories [25]. Through engineering or doping, the durability of the device can be improved, similar to what was observed for amorphous carbon storage devices [26]. Retention of nonvolatile states tested up to a week (**Figure 22(c)**) is already sufficient for certain neuromorphic applications with short-term and medium-term plasticity [27]. The subnanometric thickness of monolayers is promising for the realization

of ultrahigh densities. With a free step of 10 nm, the atomic density of 1015/mm3

single-bit single-level storage device, this corresponds to a theoretical surface density of

Modern switches are implemented using transistor or microelectromechanical devices, both of which are volatile, and the latter also requires a large switching voltage that is not suitable for mobile technologies. Recently, phase change switches have attracted interest [29], but the

provide sufficient space to simulate the density of human synapses (~10<sup>9</sup>

 **memristors**

would

) [28]. For a

/mm3

monolayer operating at 50 GHz.

strated using a MoS2

until the next cycle [23].

**Figure 21.** Scheme of a TMD sandwich based on MoS2

84 Advances in Memristor Neural Networks – Modeling and Applications

of bipolar resistive switching in a MoS2

6.4 Tbit/inch<sup>2</sup>

.

**6.2. High-frequency 2D MoS2**

**Figure 22.** Characteristics of the atomistor. (a, b) Resistance spread of MoS2 crossbar MIM devices for 150 manual dc switching cycles. (c) Time-dependent measurements of the MoS2 switch with stable storage of information for a week at room temperature. Resistance of HRS and LRS is determined by measuring the current at a small bias voltage of 0.1 V. The area of this transverse device 2 L-MoS2 is 2 × 2 μm<sup>2</sup> . (d) Experimental, nonvolatile RF switches based on a 1 × 1 μm<sup>2</sup> MoS2 monolayer show promising characteristics with an insertion loss of ~1 dB and isolation >12 dB up to 50 GHz. The cut-off frequency is ~1.8 THz. (e) Stable resistance of states with high resistance and low resistance after 1000 cycles of bending at 1% strain [23].
