2023 1446 Moesm1 Esm

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 48

nature nanotechnology

Article https://doi.org/10.1038/s41565-023-01446-8

Reconfigurable, non-volatile neuromorphic


photovoltaics
In the format provided by the
authors and unedited
Supplementary Information

1
Table of contents

Figure S1 | Thickness characterization of MoS2

Figure S2 | TEM and Raman characteristics of pristine and plasma-treated MoS2

Figure S3 | Raman spectra of MoS2 samples treated by different plasma methods

Figure S4 | STEM images and EDS mapping of the samples by different plasma treatment

methods

Figure S5 | WDS mapping and elemental spectra of sulfur and oxygen in the plasma-

treated MoS2 M/S/M devices

Figure S6 | WDS characteristics of plasma-treated MoS2

Figure S7 | XPS characteristics of plasma-treated MoS2

Figure S8 | Short-circuit current characteristics of the MSM devices at 830 nm

Figure S9 | Short-circuit current characteristics of the plasma-treated WS2 devices

Figure S10 | Photocurrent characteristics of pristine MoS2

Figure S11 | Photocurrent characteristics of the MoS2 devices treated by different plasma

methods

Figure S12 | The S/Mo ratio from STEM-EDS characterization along the MoS2 surface and

the corresponding STEM images

Figure S13 | Optical images of plasma-treated and pristine MoS2 for photocurrent mapping

Figure S14 | Photocurrent mapping of plasma-treated MoS2

Figure S15 | Photocurrent characteristics of the MSM devices with applied different voltage

pulses

Figure S16 | Output characteristics of 10 samples with different pulsing voltages

Figure S17 | Diagrams of migration path of S vacancy in atom scale

Figure S18 | Output characteristics of large short-circuit photocurrent and open-circuit

voltage

Figure S19 | Photocurrent characteristics of the M/S/M device

Figure S20 | AFM morphology and corresponding KPFM images measured from pristine

MoS2 M/S/M devices

Table 1 | Device parameters for temperature simulation in TCAD

2
Figure S21 | Structure and temperature distribution of the metal/MoS2/metal device for

TCAD temperature simulation and device current during voltage programming

Table 2 | Device parameters for TCAD simulations

Figure S22 | Distribution of carrier doping and work function in TCAD model

Figure S23 | The 11 photoresponse states are measured at 520 nm and zero bias of Vds

Figure S24 | Histogram statistics of 11 photoresponse states from Figure S23

Figure S25 | I-t characteristic of plasma-treated MoS2 device for short-circuit photocurrent

Table 3 | Benchmark the performance of the neuromorphic photovoltaic devices

Figure S26 | Photocurrent characteristics of the MoS2 devices programmed and tested in

vacuum and atmospheric conditions

Figure S27 | Multiple photoresponse states used for Image processing by a 3 × 3 array

Figure S28 | Use the photoresponse states for MINIST recognition

Figure S29 | Conductance modulation with multiple voltage pulses in dark

Figure S30 | Training strategy for object detection by using optical and electronic states

Figure S31 | The training and testing set

Figure S32 | Schematic diagram of heat map processing using the memristor array

3
a b

Figure S1. Thickness characterization of MoS2. a, AFM topography image of MoS2

flakes obtained by a mechanical exfoliation method. b, Height profiles taken along the

horizontal red lines in panel (a). The thickness of the MoS2 flake is around 10 nm.

4
a b c

① ②


10 μm 2 nm

Figure S2. TEM and Raman characteristics of pristine and plasma-treated MoS2.

Optical microscopy image (a) and transmission electron microscope (TEM) image (b) of

MoS2 flakes before (②) and after (①) plasma treatment. c, Raman spectrum of MoS2

flakes before (②) and after (①) plasma treatment, measured using an excitation

wavelength of 520 nm.

The thinning down of the MoS2 flake after O2/Ar plasma treatment shows different

thickness/contrast compare to pristine MoS2 which is confirmed by the optical/TEM images

shown in Fig. S2a and S2b. Fig. S2c shows the two Raman features of the MoS2, which
1
correspond to 𝐸𝐸2𝑔𝑔 and 𝐴𝐴1𝑔𝑔 peak, before and after O2/Ar plasma treatment. The red shift of
1
𝐸𝐸2𝑔𝑔 peak and the blue shift of 𝐴𝐴1𝑔𝑔 peak as well as the decreasing in intensities of these

two peaks further confirm the thinning down of the MoS2 flake after plasma treatment1, 2.

5
b

Figure S3. Raman spectra of MoS2 samples treated by different plasma methods. a,

The lattice vibration modes of MoO3 from 200 to 320 cm-1. b, The lattice vibration modes

of MoS2 from 350 to 430 cm-1. The black, red, blue and green curves represent the Raman

spectra for pristine MoS2, MoS2 treated by pure O2 plasma, MoS2 treated by O2/Ar plasma

in sequence, and MoS2 treated by pure Ar plasma, respectively.

To further confirm the lattice distortions caused by plasma treatment and to understand

their formation, Figure S3 shows the Raman spectra of MoS2 treated by different plasma.

The oxidation-induced peak at 225 cm−1 (red line, Figure S3a) indicates the formation of

defected MoO3-disordered domains and Mo-O bonds in the MoS2 layer3, 4, after pure O2

plasma treatment5-7. However, they disappeared when the samples were further treated

by Ar plasma (blue line, Figure S3a). The Raman characterizations demonstrated that the

O2 plasma can introduce α-MoO3 in MoS2, however, they can be fully removed by following

Ar-plasma treatment. Figure S3b shows the typical Raman spectra of MoS2 crystals which
1
correspond to 𝐸𝐸2𝑔𝑔 and 𝐴𝐴1𝑔𝑔 peak.

6
a
Pristine Oxygen Molybdenum Sulfur

MoS2

SiO2

b
O2 Plasma Oxygen Molybdenum Sulfur

MoS2

SiO2

c
O2/Ar Plasma Oxygen Molybdenum Sulfur

MoS2

SiO2

d
Ar Plasma Oxygen Molybdenum Sulfur

MoS2

SiO2

Figure S4. STEM images and EDS mapping of the samples by different plasma

treatment methods. Red: oxygen, purple: molybdenum, green: sulfur a, Pristine MoS2. b,

O2 plasma treatment for 10 s. c, O2/Ar plasma treatment for 10 s/30 s. d, Ar plasma

treatment for 30 s. Scale bar, 5 nm.

Note that the pristine MoS2 without plasma treatment showed negligible oxygen

elements, as shown in Figure S4a. This is reasonable because the self-passivated pristine

MoS2 is free of dangling bonds which makes it hard to interact with O2 or H2O in the

atmosphere. However, after plasma treatment, the MoS2 flake presents many sulfur

vacancies (dangling bonds) and is slightly oxidized when the sample is exposed to the

7
atmosphere. This is inevitable during venting the chamber of plasma equipment or

microfabrication process.

Therefore, we did more STEM-EDS as shown in Figure S4 to demonstrate the

oxidation degree of the MoS2 flakes treated by different plasma (O2/Ar plasma in sequence,

pure O2 plasma, and pure Ar plasma). Specifically, Figure S4a shows the STEM-EDS

mapping of different elements in pristine MoS2 without plasma treatment, and the other

samples were treated by (Figure S4b) pure O2 plasma for 10 s, (Figure S4c) O2/Ar plasma

for 10 s/30 s in sequence, and (Figure S4d) pure Ar plasma for 30 s, respectively.

After being treated by pure O2 plasma for 10 s, a very strong O signal was observed,

suggesting that the oxygen has replaced the sulfur in the MoS2 lattice through an oxide-

forming reaction8, 9 (Figure S4b). The formation of MoO2 and MoSO are not preferable in

O2 plasma-treated MoS2 which has been confirmed by previous reports4, 10. Such oxidation

belongs to the growing lattice distortion engendered by the presence of MoO3 defect sites

created by O bombardment2, 4 (2MoS2 + 7O2 → 2MoO3 + 4SO2).

Indeed, the defected MoO3-disordered domains formed during O2 plasma treatment.

However, further treatment by Ar plasma significantly removed the O elements (Figure

S4c). This process shifts from oxygen insertion into the lattice to the Ar-plasma-dominated

removal of oxidized species in the lattice8, 11. We believe that the O signal comes from the

oxidation of an unstable Mo-terminated surface which is similar to the sample treated by

pure Ar plasma (Figure S4d).

8
a b c
(i) −V (i) +V (i)

Drain Channel Source Drain Channel Source Drain Channel Source

O O O

(ii) (ii) (ii)

d e

Large change
No change

Figure S5. WDS mapping and elemental spectra of sulfur and oxygen in the plasma-

treated MoS2 M/S/M devices. a, The WDS mapping of oxygen atoms distribution (i), and

elemental spectrum (ii) of plasma-treated MoS2 devices without programming. b, The WDS

mapping of oxygen atoms distribution (i), and elemental spectrum (ii) of plasma-treated

MoS2 devices with negative voltage programming. c, The WDS mapping of oxygen atoms

distribution (i), and elemental spectrum (ii) of plasma-treated MoS2 devices with positive

voltage programming in sequence. Scale bar, 400 nm. d, the intensity of sulfur elements

along the channel with/without voltage programming. e, the intensity of oxygen elements

along the channel with/without voltage programming.

The plasma-treated MoS2 will be slightly oxidized during the device fabrication process.

To further characterize the migration of oxygen and sulfur vacancies driven by the electric

field, we did more WDS measurements with new devices, as shown in Figure S5. All the

samples for WDS measurements still follow our standard recipe with O2/Ar plasma

treatment in sequence (10 s/30 s). We use a specific LED1L crystal in EPMA-WDS which

allows extremely sensitive detection of oxygen, this crystal enables oxygen signal detection

with an ultralow probe current of 10-10 A which effectively reduces the charging effect on

9
the devices.

As shown in Figure S5a-i, the metal/MoS2/metal device without electric programming

showed uniform distribution of oxygen elements. After applying a negative voltage pulsing

(amplitude, -10 V; duration, 10 s, Figure S5b-ii), the distribution of oxygen atoms is very

similar to the initial state without voltage programming (Figure S5a-i). Figure S5b-ii shows

that the migration of oxygen is not observed along the channel after voltage programming.

To demonstrate the reversible migration, positive (amplitude, +10 V; duration, 10 s)

electric pulses are applied in sequence, as shown in Figure S5c. And, we still did not

observe oxygen migration along the channel after positive programming in sequence

(Figure S5c).

Figure S5d shows the sulfur intensity as a function of channel length before (dashed

black line) and after (dashed red line) bias programming. As can be seen, after bias

programming, the channel shows obvious sulfur gradient concentrations demonstrating the

migration of sulfur vacancies. However, in Figure S5e, the oxygen intensity of MoS2 after

bias programming shows negligible changes along the channel compared to the initial state.

It further confirms that there is no oxygen migration in our devices even though there exist

oxygen elements at the metal/MoS2 junction regions.

10
a b c

MoS2 Channel

Figure S6. WDS characteristics of plasma-treated MoS2. a, Scanning electron

microscopy (SEM) image of the plasma-treated metal/MoS2/metal device, the region

between the blue line is MoS2 Channel. Wavelength dispersive X-ray spectroscopy (WDS)

of S-atom (b) and Mo-atom (c) distribution below the metal electrode and MoS2 channel

regions.

The energy dispersive spectroscopy (EDS) can hardly estimate the contents of S and

Mo elements precisely which attribute to the spectral overlap between the S Kα (2.30 keV)

and the Mo Lα (2.29 keV) X-ray emission lines. We carried out the WDS measurement to

estimate the contents of S and Mo elements quantitatively12, 13. Obviously, the content of

S is much higher than that of Mo for the pristine MoS2 channel as shown in Fig. S6b and

S6c. The ratio of S/Mo is about 1.95 which is close to the stoichiometry of the S and Mo

elements in pristine MoS2. On the contrary, the S signal below the electrodes is much

weaker than the S signal in the channel with a ratio of S/Mo of about 1.42. The different

contrast ratio of S/Mo between pristine MoS2 and plasma-treated MoS2 suggests that S

vacancies appear under the electrode. This result is consistent with the experimental

process, that MoS2 under the electrode was treated by plasma.

11
a b

Pristine
Plasma treated

S 2s

Figure S7. XPS characteristics of plasma-treated MoS2. a, XPS spectra for the pristine

and plasma-treated MoS2 samples. b, S and Mo atomic ratio for pristine and plasma-

treated MoS2 samples.

The XPS spectra of Mo 3d and S 2p for MoS2 film showed the relative content of the

elements. And by calculating the atomic percent of S/Mo, the S/Mo atomic ratio of pristine

MoS2 is 1.94 (nearly to theoretical value: 2). The plasma-treated MoS2 atomic ratio is 1.5

showing the loss of sulfur atoms during plasma treatment11.

12
a b c

Figure S8. Short-circuit current characteristics of the MSM devices at 830 nm. I-V

curves for positive (a) and negative (b) short-circuit photocurrent under 830 nm illumination.

c, I-P curves for positive and negative short-circuit photocurrent at 830 nm with different

power densities.

The pulsing voltage applied to the metal/MoS2/metal device can change the sign and

magnitude of short-circuit photocurrent (Isc). The photocurrent is about 0.03 nA or -0.03 nA

at 830 nm light illumination with Vds= 0 V. It demonstrates the possibility of achieving

positive and negative photoresponse in the near-infrared wavelength region.

13
a b c

Au Au

WS2

Figure S9. Short-circuit current characteristics of the plasma-treated WS2 devices.

a, Optical microscopy image of the plasma-treated WS2 device. Scare bar, 20 μm. b,

Negative short-circuit photocurrent (Isc=-69 nA) under 520 nm illumination. c, I-V curves of

the plasma-treated WS2 device after programming with negative voltage showing that the

short-circuit photocurrent is tuned to a positive value of 332 nA. Therefore, the plasma-

treated WS2 showed the same tunable photocurrent behavior as the plasma-treated MoS2.

14
4.2x10-9 Pristine MoS2
Reading

2.1x10-9

Current (A)
0.0

-1.8x10-3

-1.9x10-3
Vds=0 V Writing
-3
-2.0x10
λ=520 nm

70 80 90 100 110

Time (s)

Figure S10. Photocurrent characteristics of pristine MoS2. Photoresponse of pristine

MoS2 without plasma treatment before and after 15 V/10 s programming ( Vds=0 V, λ=520

nm).

Compare to plasma-treated MoS2 devices, the photocurrent of pristine MoS2 did not

show observable change even programmed by a large voltage pulse (15 V/10 s). This is

because the pristine MoS2 crystal shows negligible native sulfur vacancies. And almost no

sulfur vacancies migrate by applying electric fields.

15
a b c

Large change

No change

Minor change

Figure S11. Photocurrent characteristics of the MoS2 devices treated by different

plasma methods. a, Photocurrent characteristics of the device treated by pure Ar plasma

for 40 s; the Isc is changed by 20 nA after 800 cycles of voltage programming (10 V, 100

ms). b, Photocurrent characteristics of the device treated by O2/Ar plasma 10 s/30 s in

sequence; the Isc is changed by 145 nA after 800 cycles of voltage programming (10 V, 100

ms). c, Photocurrent characteristics of the device treated by O2 plasma for 40 s; the Isc is

changed by 1.5 nA after 800 cycles of voltage programming (10 V, 100 ms).

The photoresponse characteristics further confirm that the migration of S vacancies

instead of O elements is the dominant mechanism, as shown in Figure S11. For the pristine

metal/MoS2/metal photovoltaic detectors, they did not show short-circuit photocurrent after

hundreds of voltage programming tests (Figure 2d in the main text and Figure S10). When

small amounts of S vacancies are introduced by pure Ar plasma treatment, the sign and

magnitude of the device’s short-circuit photocurrent are still hard to tune (Figure S11a).

When large amounts of S vacancies are formed not only at the surface but also inside the

MoS2 lattice (treated by O2/Ar plasma in sequence), the metal/MoS2/metal photovoltaic

detectors show a large and stable reconfigurable short-circuit photocurrent (tunable

sign/magnitude, Figure S11b). When the sample was treated by pure O2 plasma, the

devices showed negligible reconfigurable short-circuit photocurrent even 800 cycles of

voltage programming are applied (Figure S11c).

16
MoS2
Inside Surface
Pristine a

Ar Plasma b

O2/Ar Plasma c

Figure S12. The S/Mo ratio from STEM-EDS characterization along the MoS2 surface

and the corresponding STEM images. a, Pristine MoS2 (blue line). b, MoS2 with pure Ar

plasma treatment (red line). c, MoS2 with O2/Ar plasma treatment in sequence (black line).

To quantitatively analyze the elements of plasma-treated MoS2, the S/Mo ratio in the

vertical direction of the MoS2 surface is characterized by the STEM-EDS method. The ratio

of S/Mo keeps the same for pristine MoS2 without plasma treatment (blue line, Figure

S12a). However, for pure Ar plasma-treated MoS2, the ratio of S/Mo decreases near the

MoS2 surface (red line and corresponding STEM image, Figure S12b). This is because

the formation energy of the sulfur vacancies is the lowest among all types of defects3.

More interestingly, the MoS2 treated by O2/Ar plasma in sequence not only shows high

sulfur vacancies but also enables sulfur vacancies formation inside the MoS2 lattice (black

line and corresponding STEM image, Figure S12c). It is very different from the sample

treated by pure Ar-plasma in which the sulfur vacancies can only be found in the

amorphous region of the MoS2 surface. This is because the O2 plasma could make

energetic oxygens not only interact with the surface atoms but also propagate deeply inside

the lattice to oxidize MoS25. After that, the following Ar plasma will remove O and S

elements leaving sulfur vacancies in the lattice10.

17
Figure S13. Optical images of plasma-treated and pristine MoS2 for photocurrent

mapping. The optical microscopy images of plasma-treated (a) and pristine MoS2 (b)

devices for photocurrent mapping measurement. The white dashed lines indicate the MoS2

and source/drain electrodes.

18
a

Source

Cr/Au
MoS2

Cr/Au
(−)
Drain 4 μm

Figure S14. Photocurrent mapping of plasma-treated MoS2. a, Photocurrent mapping

from plasma-treated MoS2 with applied -15 V/15 s pulse voltage following the state in

Figure 2a-iii in the main text. Vds=0 V, λ=520 nm. b, The distribution of photocurrent is

taken along the horizontal black line, the red and blue dashed lines show the position of

photocurrent close to the metal electrodes.

19
a b c

Figure S15. Photocurrent characteristics of the MSM devices with applied different

voltage pulses. a, Device output characteristics programmed by different voltage pulses

(10 V 100 ms). The wavelength is λ=520 nm. b, I-V curves extracted from the yellow region

in panel (a) with applied cycles of 0, 3, 4, 6 and 8 voltage pulses. c, The open-circuit voltage

extracted from (a) under the different cycles of pulsing voltage. Also, by optimizing the

cycles of pulsing voltages, the open-circuit voltages can be changed from negative to

positive.

20
a b c

d e f

g h i

Figure S16. Output characteristics of 10 samples with different pulsing voltages. (a)-

(e), Variation of short-circuit photocurrent (Isc) changes from positive to negative with

applied negative pulsing voltage. Scale bar, 8 μm. (f)-(j), Variation of short-circuit

photocurrent (Isc) changes from negative to positive with applied positive pulsing voltage.

Scale bar, 8 μm.

Here, we chose 10 different metal/MoS2/metal devices with the thickness of MoS2

changing from 10 to 50 nm and the channel of devices is around 1-2 μm. When the

negative pulsing voltages are applied, the short-circuit photocurrent changes from positive

to negative as shown in Figures S16a-S16e. When the positive pulsing voltages are

applied, the short-circuit photocurrent changes from negative to positive as shown in

Figures S16f-S16j.

21
Figure S17. Diagrams of migration path of S vacancy in atom scale. a-e, Migration

process of sulfur vacancy in the lattice, initial state (a) to final state (e). The relevant bond

lengths of Mo-S atoms are marked. f, Migration barrier of the movement of S vacancy.

To understand the migration process of S vacancies, we performed the first-principles

calculations based on the density functional theory to describe the migration path of S

vacancy from the atomic scale. To consider S vacancy, the 3 × 3 × 1 supercell of monolayer

MoS2 is constructed. The migration of S vacancy can be regarded as the nearest-neighbor

S atom occupying the original site of S vacancy. The migration path of S vacancy is

22
presented in Figure S17. Structures in Figure S17a and Figure S17e are selected as the

initial and the final state, respectively. Figure S17c corresponds to the transition state and

the bond length of Mo-S is 2.684 Å, which is 11.9% longer than that of the pristine one in

Figure S17a. As the migration continues, the bonding character of Mo and S atoms has

changed. Therefore, the migration process of S vacancy is accompanied by the breaking

of old bonds (de-bonding) and the formation of new bonds (bonding). The calculated barrier

for migration of S vacancy is 2.145 eV, as shown in Figure S17f. In our study, the applied

electric field (105 V/cm) can let the S vacancies obtain enough energy to overcome the

barrier and migrate in the devices14.

Calculation details: The first-principles calculations were carried out with Vienna Ab-

initio Simulation Package (VASP)15, 16 using the projector augmented wave (PAW)17

method. The exchange-correlation interactions were treated using the Perdew-Burke-

Ernzerhof (PBE)18 functionality. The cutoff energy was set to 400 eV. The energy

convergence criterion and the forces on atoms were set to 10-5 eV and 0.02 eV/Å,

respectively. The 3 × 3 × 1 supercell of monolayer MoS2 was built for considering S vacancy.

The k-mesh was set to 3 × 3 × 1 during structural optimization. A vacuum layer of 15 Å was

added in the vertical direction to avoid the interactions of neighboring sheets. The migration

path of S vacancy to the nearest-neighbor S atom was calculated using the nudged-elastic-

band (NEB) method19 with 3 images.

23
a b

Figure S18. Output characteristics of large short-circuit photocurrent and open-

circuit voltage. a, The optical microscopy image of the metal/MoS2/metal device. Scale

bar, 4 μm. b, I-V curve under dark and illumination conditions, λ=520 nm. The two-terminal

device shows an ultrahigh Isc of -795 nA and Voc of 60 mV. The device's area is 13.6 μm2,

also light spot diameter is 200 μm and power of the laser is 4.98 mW. Therefore, the

calculated responsivity is 369.2 mA/W.

24
a b

Figure S19. Photocurrent characteristics of the M/S/M device. (a) Optical microscope

image of the MoS2 M/S/M device with O2/Ar plasma treatment. Scale bar, 8 µm. (b)-(c) Isc

as a function of optical power density showing a highly linear relationship from 16-2600

mW/cm2 (b) and 16-1000 mW/cm2 (c).

To demonstrate the sensitivity of the plasma-treated metal/MoS2/metal neuromorphic

photovoltaic devices, we fabricated new devices to test the photocurrent response under

weak light illumination. As shown in Figure S19, the devices show almost linear short-

circuit photocurrent under optical power changing from 16 mW/cm2 to 2600 mW/cm2.

Therefore, the devices can detect illumination level which is almost 10X lower than that of

solar irradiance (~100 mW/cm2).

25
a b c

1
2
3
3
1

Figure S20. AFM morphology and corresponding KPFM images measured from

pristine MoS2 M/S/M devices: (a) and (b). The magnitude of the electric potential was

extracted from the KPFM images. c, The potential difference between the MoS2 channel

and Au electrodes. And points 1,2,3 in (c) are consistent with points in (b).

The variation of potential is between the metal/MoS2 interface and the MoS2 channel.

As can be seen, the potential is symmetrically distributed on both sides of the metal/MoS2

junctions, and there is no potential change along the region of the MoS2 channel. It

demonstrates that, for pristine MoS2, the electric field is symmetrically distributed.

26
Table 1. Device parameters for temperature simulation in TCAD
MoS2 Si SiO2

16 -3 20 -3
Doping concentrations 10 cm 10 cm -

Electron affinity 4.20 eV 20 4.07 eV 0.9 eV

Bandgap 1.30 eV 20 1.12 eV 9 eV

Thermal conductivity 0.52 W/(K• cm) 21 1.70 W/(K• cm) 22 0.014 W/(K• cm) 23

3 3 3
Heat Capacity 1.89 J/(K• cm ) 24 1.63 J/(K• cm ) 23 1.67 J/(K• cm ) 25

Self-heating simulation method: In addition to solving sets of coupled differential

equations namely Poisson’s equation, current continuity equations for electrons and

temperature-related equations should be also considered for self-heating simulation. And

several models are also considered in TCAD simulation, for example, carrier

recombination/generation model, mobility model, thermodynamic model, and thermal

conductivity model. Furthermore, thermal boundary conditions must be applied to solve the

temperature equations. At thermally conducting interfaces, thermally resistive boundary

conditions will be considered, which characterize the thermal contact between the

semiconductor and adjacent material. The solution step in the basic quasi-static solution

process is voltage, while the transient method with a solution time step is required to

consider the effect of pulse voltage on temperature.

27
a Drain Source cm-3 b cm-3
1020 MoS2 (10 nm) 1020
Si (1 µm) SiO2 (300 nm)
1016 1016
c d
10V 0V K -10V 0V K
370 370

300 300
e K f K

300 300

g h

i j

Figure S21. Structure and temperature distribution of the metal/MoS2/metal device

for TCAD temperature simulation and device current during voltage programming.

a, Structure of the MoS2 M/S/M device on SiO2/Si substrate for TCAD simulation. Scale

bar, 1 μm. b, Zoomed-in structure of the MoS2 M/S/M device on SiO2/Si substrate. Scale

bar, 100 nm. c, Distribution of temperature in the MoS2 M/S/M device on SiO2/Si substrate

with 10 V/100 ms pulsing voltage. Scale bar, 1 μm. d, Distribution of temperature in the

MoS2 M/S/M device on SiO2/Si substrate with -10 V/100 ms pulsing voltage. Scale bar, 1

μm. e, Temperature distribution of the MoS2 M/S/M device after switching off pulsing

voltage (10 V) for 100 ms (cooling down process). Scale bar, 1 μm. f, Temperature

distribution of the MoS2 M/S/M device after switching off pulsing voltage (-10 V) for 100 ms
28
(cooling down process). Scale bar, 1 μm. g, Device current with 10 V/100 ms bias in the

experiments. h, V-t and I-t characteristics with 10 V/100 ms bias in the TCAD simulation. i,

j, Temperature distribution of horizontal direction in MoS2 of (c) and (d).

We performed Sentaurus-TCAD thermal simulation of metal/MoS2/metal device on

SiO2/Si substrate with a channel length of 1µm and metal electrode width of 4 µm, as

shown in Figure S21a and Figure S21b. The thicknesses of MoS2 and SiO2 are 10 nm

and 300 nm. The heat-related coefficients such as thermal conductivity, heat capacity and

related energy band parameters are shown in Table 1 and Table 2. We applied 10 V/100

ms and -10 V/100 ms pulsing voltages to the metal/MoS2/metal devices. And the initial

temperature of the device is set to 300 K. The current in the TCAD simulation is set to 50-

100 µA which is consistent with the experimental results (50-100 µA), as shown in Figure

S21g and Figure S21h.

As shown in Figure S21, with 10 V/100 ms pulsing voltage, the temperature of the

metal/MoS2 junction region is increased to 370 K. However, the heat rapidly diffused in the

vertical and horizontal directions by the metal electrodes, SiO2/Si substrates, and air. Note

that the temperature (<100 °C) induced by programming voltage is significantly lower than

800 °C which is the critical temperature for the diffusion of sulfur vacancy26. It demonstrates

that the Joule heating effect is negligible for the diffusion of sulfur vacancies. Therefore,

the migration of sulfur vacancies is mainly driven by applied electric field. In our

experiments, almost no devices were burned with applied 10 V/100 ms or -10 V/100 ms

pulsing voltages further demonstrating that the self-heating is negligible. Figure S21 also

compares the device temperatures under +10 V (forward bias) and -10 V (reverse bias)

programming condition showing the same temperature increment but the opposite position

of the metal/MoS2/metal device.

29
Table 2. Device parameters for TCAD simulations
Device parameters
18 16 -3
Doping concentration (NC) 10 -10 cm
Electron affinity (χ) 4.2 eV
Band gap (Eg) 1.3 eV
Channel length (L) 1 μm

Schottky work function (SW) 5.1-4.98 eV (step 0.01)

The device parameters for the simulation are detailed in Table 2. According to the

experiment, by using the module called SDE in Sentaurus Workbench to design the

structure. The 1 μm channel was built and two electrodes contact were constructed at both

sides of the channel. The parameter file in SDE was revised for the MoS2 material. Also,

the optoelectronic characterization of the device was performed in the SDEVICE module.

Two electrodes contact were set for the Schottky contact by defining the Schottky work

function in the codes. The doping in MoS2 varies at 1016-1018 cm-3. And, in the physics

section, the recombination and non-local tunneling model was taken into consideration,

T=300 K. In addition, the light section was also introduced in the device, and the

wavelength, intensity, and direction of light were important for the simulation. Finally, the

electron and hole current got from the solution of the Poisson equation with different sweep

voltage ranges. Simulation results were all got from the Svisual module. In the simulation,

combined the result of KPFM and WDS mapping. The sulfur vacancies can change the

Schottky barrier beside the two electrode contacts and also change the distribution of

doping. So, we designed the variable Schottky work function on one side of the electrode

contact and consistent with the variation of concentration to see the changes in short-circuit

current of the devices.

30
Figure S22. Distribution of carrier doping and work function in TCAD model. a-(1),

(3), (5), (7), (9), (11), The barrier height varies with the concentration distribution at the

source and drain, and the maximum gradient of doping is 1018-1016 cm-3 with 4.98-5.1 eV

metal work function at (11). a-(2), (4), (6), (8), (10), (12), Electron density distribution under

illumination at 0V bias at (1), (3), (5), (7), (9), (11). b, TCAD-simulated photocurrent

characteristics of the plasma-treated MoS2 devices with shallow-energy level and deep-

energy level traps. The concentration of shallow/deep level traps is 1014 cm-3. And the work

function is 5.05 eV at source/drain in the MoS2 M/S/M photovoltaic detector.

To accurately analyze photocurrent modulation by the Schottky barrier and the non-

31
uniform distribution of S vacancies, we considered the non-uniform carrier concentration

induced by gradient S vacancies and recombination centers in the TCAD simulation. We

further optimized the device parameters and structures with a MoS2 thickness of 10 nm

and MoS2 channel of 1 µm which are consistent with the physical devices. On one hand,

the S vacancies can reduce the Schottky barrier height. On the other hand, the S vacancies

can increase the carrier concentration27. Therefore, in the TCAD simulation, we reduced

(or increased) the Schottky barrier height while increasing (or reducing) the carrier

concentration. If the work function is 5.1 eV for both source and drain contact, the MoS2

channel is set to be uniform carrier concentration of 1016 cm-3. And when the work function

is set to 4.98 eV (or 5.1 eV), the corresponding carrier concentration is set to 1018 cm-3 (or

1016 cm-3). By changing the Schottky barrier and carrier concentration simultaneously, we

still observed the tunable sign/magnitude of short-circuit photocurrent in the TCAD

simulation, which agrees well with our experimental results as shown in Figures 3f and

2g (in the main text). And the corresponding carrier concentration, potential barrier, and

electron current density distributions under illumination are all shown in Figure S22a.

To further check the effect of recombination centers on photocurrent, both deep-

energy-level defects and shallow-energy-level defects are considered in the Sentaurus-

TCAD simulation. The shallow/deep-energy band is 0.1 eV/0.5 eV below the conduction

band minimum28, 29. The trap concentration is set to 1014 cm-3 which is higher than 1012 cm-
3~1013 cm-3 in the pristine MoS2. As shown in Figure S22b, after introducing shallow/deep-

energy-level defects, the sign/magnitude of photocurrent can still be modulated. For

shallow-energy-level defects, no significant change in short-circuit current compared to

pristine MoS2. But for deep-energy-level defects, the short-circuit current was reduced.

This is because the defect-induced recombination centers can suppress the photovoltaic

effect30. However, even the defect-induced recombination centers are artificially introduced

in the TCAD simulator, the metal/MoS2/metal devices still show tunable sign/magnitude of

short-circuit photocurrent.

32
a b c

d e f

g h i

j k

Figure S23. The 11 photoresponse states are measured at 520 nm and zero bias of

Vds. a-e, Positive photoresponse states obtained for different cycles of negative pulsing

voltages (-10 V/100 ms). f, Nearly zero photoresponse state achieved from the positive

pulsing voltage (10 V/100 ms with 50 cycles of pulsing voltage) after state 5 in Figure S23e.

g-k, Negative photoresponse states from different cycles of positive pulsing voltages (10

V/100 ms).

33
Figure S24. Histogram statistics of 11 photoresponse states from Figure S23.

By counting the magnitude of the photocurrent from Figure S23, the average value

was calculated. The different values of short-circuit photocurrent are 2 nA, 4 nA, 5.5 nA, 9

nA, 11 nA, 0 nA,-1.5 nA, -3.5 nA, -5 nA, -7 nA, -10 nA. Considering the device area and

the size of the laser spot, 11 responsivity states were calculated from different Isc, 6.3 mA/W,

12.0 mA/W, 16.5 mA/W, 27.0 mA/W, 33.0 mA/W, 0.0 mA/W, -4.5 mA/W, -10.5 mA/W, -15.9

mA/W, -21 mA/W, -29.7 mA/W. Large distinguishability of each photoresponse state and

repeatable modulation can be achieved. And these positive and negative states can be

useful for neural network computing. The photoresponse is related to the light intensity and

photocurrent and can be calculated by the formula as shown in equation (1):

𝐼𝐼
𝑅𝑅 = (1)
𝑃𝑃
Considering the light spot area is larger than the device area, the result should be

normalization as shown in equation (2):

𝐼𝐼 ∙ 𝑆𝑆𝑙𝑙𝑙𝑙𝑙𝑙ℎ𝑡𝑡
𝑅𝑅 = (2)
𝑃𝑃 ∙ 𝑆𝑆𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
In the experiment, the light spot radius is 200 μm and the device area is 84 μm2. The

light intensity is 4.98 mW.

34
6.0x10-9
5.0x10-9 Retention time > 1000s
-9
4.0x10
3.0x10-9
2.0x10-9
Isc (A)

1.0x10-9
0.0

-1.0x10-5
-1.1x10-5
300 450 600 750 900 1050
Time (s)
Figure S25. I-t characteristic of plasma-treated MoS2 device for short-circuit

photocurrent. The retention time of the photoresponse state is large than 1000 s under

520 nm illumination at 0V bias.

35
Table 3. Benchmark the performance of the neuromorphic photovoltaic devices

Architecture Retention Photoresponsivity Linear of Isc-P Multi-states


(Maximum) (Maximum) (Photoresponsivity)

This work Two-terminal >1000 s 369.2 mA/W 2600 mW/cm


2 11

Thomas et al 31 Four-terminal Volatile 60 mA/W 100 mW/cm


2 -

Li et al 32 Three-terminal 2000 s 60 mA/W - 2

Zhai et al 33 Three-terminal Volatile 350 mA/W 40 mW/cm


2 6

Liu et al 34 Two-terminal >800 s 0.22 mA/W 150 mW/cm


2 25

Hu et al 35 Two-terminal >3000 s 98.8 mA/W - 4

36
a b

Air Vacuum

Figure S26. Photocurrent characteristics of the MoS2 devices programmed and

tested in vacuum and atmospheric conditions. a, Photocurrent characteristics of the

O2/Ar plasma 10 s/30 s treated MoS2 M/S/M photovoltaic detector in the atmosphere; the

Isc switches from 130 nA to 350 nA after 500 cycles of positive voltage programming (10 V,

100 ms) and switches from 350 nA to -160 nA after 2000 cycles of negative voltage

programming (-10 V, 100 ms). b, Photocurrent characteristics of the O2/Ar plasma 10 s/30

s treated MoS2 M/S/M photovoltaic detector in the vacuum; the Isc switches from 110 nA to

660 nA after 500 cycles of positive voltage programming (10 V, 100 ms) and switches from

660 nA to -300 nA after 2000 cycles of negative voltage programming (-10 V, 100 ms).

To exclude the effect of atmospheric O2 during bias programming and photocurrent,

we carried out a photoresponse switching experiment in the atmosphere and vacuum,

respectively (Figure S26). Note that, no matter the devices are programmed and tested in

air or vacuum conditions, they all showed observable reconfigurable photoresponse

(tunable sign/magnitude). This further confirms our conclusion that the migration of S

vacancies instead of O elements is the dominant mechanism. Therefore, our

metal/MoS2/metal neuromorphic devices can work even they are packaged.

37
Figure S27. Multiple photoresponse states used for Image processing by a 3 × 3
array.

The operation of Gaussian blur is emulated by a 3 × 3 array with setting the 9 devices

to 3 different photoresponse states, and the distribution of the photoresponse states of

different devices is shown in Figure S27. The three different gray pictures (Cameraman,

38
Lena, Peppers) were used for image processing in Figure S27a. The neural network for

image processing, the single-layer perceptron was used and the formula of I = ∑ 𝑃𝑃 × 𝑅𝑅

was performed by matrix multiplication (MAC) as shown in Figure S27b. The gray value

was reflected in different power intensities, and the weights in the network used multi-

photoresponse states. The Experiment results emulated from the photoresponse states

6.3 12.0 6.3


1
matrix of the Gaussian kernel matrix �12.0 16.5 12.0� with our 3 × 3 array. The
89.7
6.3 12.0 6.3

operation of image preprocessing of the Roberts operator is emulated by a 2×2 array with

the individual device written to a specific photoresponse state, in which each cell contains

−1 0 0 −1
2 photo-memristors, different types of matrices (� � and � �) are used for edge
0 1 1 0
detection along the X-axis and Y-axis directions. The operation of image processing of the

Prewitt operator is emulated by a 3×3 array with the individual device written to a specific

−1 0 1 1 1 1
photoresponse state, different types of matrices (�−1 0 1� and � 0 0 0 � ) are
−1 0 1 −1 −1 −1

used for edge detection along the X-axis and Y-axis directions. In the Roberts and Prewitt

operators, the “-1” was from the negative photoresponse state and the “1” was from the

positive photoresponse state. For results of three different gay pictures with Gauss,

Roberts-x, Roberts-y, Prewitt-x and Prewitt-y were shown in Figure S27c, Figure S27d

and Figure S27e. Using the experimental photoresponse states, both image processing

methods can get better edge detection results.

39
a b
33.0 27.0 12.0 0.0 -10.5 -21.0
R(mA/W)

16.5 6.3 -4.5 -15.9 -29.7

•••
•••

Output
•••

Input
c d

Figure S28. Use the photoresponse states for MINIST recognition. a, Dividing the

continuous states into 11 photoresponse states, the structure schematic diagram of single

layer perceptron (SLP) also shows in (a). b, Recognition accuracy of MINIST 0-9 numbers

with float states and discrete states. The recognition accuracy of numbers 0 (c) and 1 (d)

with over 95%. Floating point weight is always used in computing for single-layer

perceptron.

The photo-responsivity matrix can also be programmed as a single-layer perception

neural network for image recognition. In this section, we trained a fully connected (FC)

layer to recognize images of handwriting numbers “0” to “9” from the MINIST data set. The

weights in FC are first trained offline with 60000 images of the training set, delivering the

recognition rate with 91% accuracy for the testing set. Then, by dividing the floating point

weight into the 11 photoresponse states (one-by-one correspondence according to the


40
scale). Compare to the recognition accuracy for single-layer perceptron with floating point

weight and 11 discrete states. The total accuracy for 0-9 images is near 90%. For image 0

and image 1, the recognition accuracy is up to 95%. So, the 11 photoresponse states are

also suitable for the recognition task of the single-layer perceptron.

41
a b

c d

Figure S29. Conductance modulation with multiple voltage pulses in dark. a,

Magnitude of voltage and duration of time for set pulse (10 V 100 ms), reset pulse (-12 V

100 ms), read pulse (2 V 100 ms). b, Current measurement with changing set pulse and

reset pulse numbers c, The magnitude of voltage and duration of time for set pulse (8 V

100 ms) and read pulse (2 V 100 ms). d, Current measurement with increasing the number

of pulse numbers.

The conductance of the plasma-treated metal/MoS2/metal can also be tunable by

applying the voltage pulses. During the set pulse and the reset pulse, a modulated

conductance state was achieved. By changing the amplitude of the voltage pulses, 168

different conductive states can be obtained.

42
Step1 Float-16 training
a float-16 weights float-16 weights

Optical part network Electronic part network


Training
Step2 Weights in optical part network discretized every 10 epochs
b
float-16 weights

Optical part network Electronic part network


Training
c Step3 Weights in electronic part network discretized every 20 epochs
weights froze
float-16 weights

Optical part network Electronic part network

Figure S30. Training strategy for object detection by using optical and electronic

states. a, Floating-16 training in step 1. b, Weights in the optical part network discretized

every 10 epochs in step 2. c, Weights in the electronic part network discretized every 20

epochs.

The training process of our object detection network is divided into three steps as

shown in Figure S30. In the first step, the weights are trained and updated with float-16

datatype. In the second step, the weights in the optical part network are discretized every

10 epochs. In the third step, the weights in the optical part network are frozen and the

weights in the electronic part network are discretized every 20 epochs.

43
Figure S31. The training and testing set.

The OSU Thermal Pedestrian Database36 is used for the evaluation of our method.

The task of the database is to conduct person detection in thermal imagery. In our

experiments, we randomly cropped patches with a size of 96×96 from the database, and

each cropped patch only contains one person. The cropped patches are randomly split

with 2131 images for training and 585 images for testing. The brightness of each pixel is

encoded into the optical power.

44
Figure S32. Schematic diagram of heat map processing using the memristor array.

a, Flowchart of the target locating method applied to the heatmap using memristor crossbar

in Spice. b, Diagram of the experimental FC demonstration with Ex situ training and weight

transfer.

Memristor has become an ideal candidate for in-sensor or near-sensor computing

device due to their nonvolatility. By converting each pixel value of the heatmap into a

voltage pulse amplitude in a specific range, and then inputting them into the memristor

array, each memristor in the crossbar array corresponds to a pixel in the heatmap, thus

576 memristors are used, and the conductance of all memristors are uniform. According to

Ohm's law and Kirchhoff's law, summation calculation is operated simultaneously both in

each row and each column. By calculating the sum of the current in each row, we compare

24 rows of current to find the maximum value, of which the row number directly indicates

the Y-axis value of the target point in the heatmap. In the same way, we obtain the X-axis

value of the target point by comparing 24 columns of current. Finally, we use the

Operational Amplifiers to transform the row and column current value to voltage value as

shown in the Figure. 4e (in the main text).

45
References

1. Li, H. et al. From Bulk to Monolayer MoS2: Evolution of Raman Scattering. Adv. Funct.
Mater. 22, 1385-1390 (2012).
2. Liu, Y. L. et al. Layer-by-Layer Thinning of MoS2 by Plasma. ACS Nano 7, 4202-4209
(2013).
3. Kim, S. H. et al. Effects of plasma treatment on surface properties of ultrathin layered MoS2.
2D Mater. 3, 035002 (2016).
4. Kang, N., Paudel, H. P., Leuenberger, M. N., Tetard, L. & Khondaker, S. I.
Photoluminescence Quenching in Single-Layer MoS2 via Oxygen Plasma Treatment. J.
Phys. Chem. C 118, 21258-21263 (2014).
5. Khondaker, S. I. & Islam, M. R. Bandgap Engineering of MoS2 Flakes via Oxygen Plasma:
A Layer Dependent Study. J. Phys. Chem. C 120, 13801-13806 (2016).
6. Pei, C. J. et al. Morphological and Spectroscopic Characterizations of Monolayer and Few-
Layer MoS2 and WSe2 Nanosheets under Oxygen Plasma Treatment with Different
Excitation Power: Implications for Modulating Electronic Properties. ACS Appl. Nano Mater.
3, 4218-4230 (2020).
7. Islam, M. R. et al. Tuning the electrical property via defect engineering of single layer MoS2
by oxygen plasma. Nanoscale 6, 10033-10039 (2014).
8. Jadwiszczak, J. et al. Oxide-mediated recovery of field-effect mobility in plasma-treated
MoS2. Sci. Adv. 4, eaao5031 (2018).
9. Nan, H. et al. Improving the electrical performance of MoS2 by mild oxygen plasma
treatment. J. Phys. D 50, 154001 (2017).
10. Brown, N. M. D., Cui, N. Y. & McKinley, A. An XPS study of the surface modification of
natural MoS2 following treatment in an RF-oxygen plasma. Appl. Surf. Sci. 134, 11-21
(1998).
11. Zhu, H. et al. Remote Plasma Oxidation and Atomic Layer Etching of MoS2. ACS Appl.
Mater. 8, 19119-19126 (2016).
12. Thomas, S. et al. Electrodeposition of MoS2 from Dichloromethane. J. Electrochem. Soc.
167, 106511 (2020).
13. Noori, Y. J. et al. Large-Area Electrodeposition of Few-Layer MoS2 on Graphene for 2D
Material Heterostructures. ACS Appl. Mater. 12, 49786-49794 (2020).
14. Li, D. et al. MoS2 Memristors Exhibiting Variable Switching Characteristics toward
Biorealistic Synaptic Emulation. ACS Nano 12, 9240-9252 (2018).
15. Kresse, G. & Furthmuller, J. Efficient iterative schemes for ab initio total-energy
calculations using a plane-wave basis set. Phys. Rev. B 54, 11169-11186 (1996).
16. Kresse, G. & Furthmuller, J. Efficiency of ab-initio total energy calculations for metals and
semiconductors using a plane-wave basis set. Comput. Mater. Sci 6, 15-50 (1996).
17. Blochl, P. E. Projector augmented-wave method. Phys. Rev. B 50, 17953-17979 (1994).
18. Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized Gradient Approximation Made Simple.
Phys. Rev. Lett. 77, 3865-3868 (1996).
19. Henkelman, G., Uberuaga, B. P. & Jonsson, H. A climbing image nudged elastic band
method for finding saddle points and minimum energy paths. J. Chem. Phys. 113, 9901-
9904 (2000).

46
20. Wu, F. et al. Vertical MoS2 transistors with sub-1-nm gate lengths. Nature 603, 259-264
(2022).
21. Sahoo, S., Gaur, A. P. S., Ahmadi, M., Guinel, M. J. F. & Katiyar, R. S. Temperature-
Dependent Raman Studies and Thermal Conductivity of Few-Layer MoS2. J. Phys. Chem.
C 117, 9042-9047 (2013).
22. Glassbrenner, C. J. & Slack, G. A. Thermal Conductivity of Silicon and Germanium
from 3oK to the Melting Point. Phys. Rev. 134, A1058-A1069 (1964).
23. Sze, S. M. Physics of Semiconductor devices (John Wiley and Sons, 1981).
24. Liu, J., Choi, G.-M. & Cahill, D. G. Measurement of the anisotropic thermal conductivity of
molybdenum disulfide by the time-resolved magneto-optic Kerr effect. J. Appl. Phys. 116,
233107 (2014).
25. Dean, D. J. Thermal design of electronic circuit boards and packages (State Mutual Book
& Periodical Service, 1985).
26. Chen, Q. et al. Ultralong 1D Vacancy Channels for Rapid Atomic Migration during 2D Void
Formation in Monolayer MoS2. Acs Nano 12, 7721-7730 (2018).
27. Lee, J. et al. Electrical role of sulfur vacancies in MoS2: Transient current approach. Appl.
Surf. Sci. 613, 155900 (2023).
28. Li, L., Long, R., Bertolini, T. & Prezhdo, O. V. Sulfur Adatom and Vacancy Accelerate
Charge Recombination in MoS2 but by Different Mechanisms: Time-Domain Ab Initio
Analysis. Nano Lett. 17, 7962-7967 (2017).
29. Yang, J., Bussolotti, F., Kawai, H. & Goh, K. E. J. Tuning the Conductivity Type in
Monolayer WS2 and MoS2 by Sulfur Vacancies. Phys. Status Solidi 14, 2000248 (2020).
30. Willardson, R. K. & Beer, A. C. Semiconductors and semimetals (Academic Press, 1977).
31. Mennel, L. et al. Ultrafast machine vision with 2D material neural network image sensors.
Nature 579, 62-66 (2020).
32. Lee, S. H., Peng, R. M., Wu, C. M. & Li, M. Programmable black phosphorus image sensor
for broadband optoelectronic edge computing. Nat. Commun. 13, 1485 (2022).
33. Pi, L. J. et al. Broadband convolutional processing using band-alignment-tunable
heterostructures. Nat. Electron. 5, 248-254 (2022).
34. Cui, B. Y. et al. Ferroelectric photosensor network: an advanced hardware solution to real-
time machine vision. Nat. Commun. 13, 1707 (2022).
35. Fu, X. et al. Graphene/MoS2−xOx/graphene photomemristor with tunable non-volatile
responsivities for neuromorphic vision processing. Light Sci. Appl. 12, 39 (2023).
36. Davis, J. W. & Keck, M. A. A two-stage template approach to person detection in thermal
imagery. In 2005 IEEE Workshop on Applications of Computer Vision (WACV) 364-369
(IEEE, 2005).

47

You might also like