3D trigonal FAPbI3‐based multilevel resistive switching nonvolatile memory for artificial neural synapse

Hybrid perovskites have attracted enormous attention in the next generation resistive switching (RS) memristor for the artificial synapses, owing to their ambipolar charge transport, long diffusion length, and tunable visible bandgap. However, the variable switch, limited reproducibility, and poor endurance are the obstacles to the practical application of the perovskite memristors. Herein, we reported a multilevel RS nonvolatile memory based on a 3D trigonal HC(NH2)2PbI3 (α‐FAPbI3) perovskite layer modified by 1‐cyanobutyl‐3‐methylimidazolium chloride ([CNBmim]Cl) and sandwiched between ITO and Au electrodes (Au/[CNBmim]Cl/α‐FAPbI3/SnO2/ITO). In contrast to the bare memristor with failure switching from low resistance state (LRS) to high resistance state (HRS), the memristor device based on the α‐FAPbI3 modified with [CNBmim]Cl (Target device) shows the retention time over 104 s with On/Off ratio (>102) and endurance up to 550 cycles. The stable RS cycle benefits from the accelerated electrons de‐trapping from the reduced defects and fast charge separation in the interface of α‐FAPbI3/electrode, leading to the rupture of conductive filaments and transition of LRS to HRS. As a two‐terminal analog synaptic device, the target device can realize random handwritten digit recognition with an impressive accuracy of 89.3% on the condition of low learning phases (500 training cycles).

Burgeoning information storage and artificial intelligence technologies are vital for settling the limitations of Moore's law and von Neumann bottleneck. 1 Resistive switching memory (ReRAM) is considered as one of the most prominent candidates in information storage and neuromorphic computing applications with powerful capabilities of simple architecture, long data retention, high switching speed, and low-power dissipation. 2,3rganic-inorganic halide perovskites (OHPs) materials have recently attracted considerable attention and made remarkable progress in the field of memristors.8][9] An early all-electrical memristor with Au/CH 3 NH 3 PbI 3−x Cl x /FTO structure was reported. 8his device showed endurance (write and erase cycles before failure) and retention (maintaining the information state even in the absence of power supply) for 100 cycles and 10 4 s, respectively.Zhu et al. 10 developed a novel memristor with the structure of a MAPbI 3 layer between two Ag/Au/Ti electrodes, and this perovskite memristor showed light-assistant synaptic behaviors.Ham et al. 11 also investigated the MAPbI 3 -based synaptic devices by simulating pattern recognition using the Modified National Institute of Standards and Technology (MNIST) database.And they achieved recognition accuracy of 82.7% after 2000 learning phases via lightassisted revision of the synaptic weights.][14][15] However, one of the major roadblocks to the development of methylammonium lead halide perovskites is the environmental and chemical stability due to the sensitivity to humidity, illumination, and thermal stress. 16In contrast, formamidinium lead iodide perovskites (HC(NH 2 ) 2 PbI 3 , FAPbI 3 ) material has better stability and a high absorption coefficient. 17,18Yang et al. 19 and Luo et al. 20 investigated the 1D yellow hexagonal FAPbI 3 (δ-FAPbI 3 ) as the active film sandwiched between Ag and Pt or Ag and FTO planar electrodes.Both found that the stable resistive switching (RS) behaviors were realized by the δ-FAPbI 3 -based devices, and the potential for neuromorphic device applications was expected.2][23] In contrast to the δ-FAPbI 3 with nonphotoactivity, FAPbI 3 with a crystal structure of 3D black trigonal FAPbI 3 (α-FAPbI 3 ) shows much better photoelectric activities, 24 which makes it more suitable for memristor to mimic the functions of the artificial visual synaptic device and to realize artificial visual neural networks.Indeed, there have been few reports so far on the α-FAPbI 3 -based memristors .The iodide vacancy (V I ˙) cluster in α-FAPbI 3 is considered so stable after forming conductive filaments (CFs) than the process of returning from the low resistance state (LRS) to high resistance state (HRS) failure, 19,20,25 which may be one of the reasons for the lack of research studies.
Herein, a ReRAM-based α-FAPbI 3 treated by 1-cyanobutyl-3-methylimidazolium chloride ([CNBmim]Cl) as the activity layer and sandwiched between Au and SnO 2 / ITO (target device, Au/[CNBmim]Cl/α-FAPbI 3 /SnO 2 /ITO structure) explored as an enabler for neuromorphic functionality.The crystal structure of the prepared α-FAPbI 3 film is confirmed by X-ray diffraction (XRD) patterns.For comparison, the control device untreated with [CNBmim] Cl (Au/α-FAPbI 3 /SnO 2 /ITO) was also assembled.With a notable difference with the control device's failure to realize the stable and multilevel RS behaviors, the target device exhibited RS behaviors with very low programming inputs for 550 cycles endurance and 10 4 s retention.A possible mechanism was also proposed to explain the RS phenomenon of the target device.Owing to the tailored interface between α-FAPbI 3 and Au by [CNBmim]Cl, the defects in this interface are passivated.As a result, the trapped electrons at the defects can be detrapped fast with the negative bias leading to the rupture of CFs and the easy process of LRS turn to HRS.What is noteworthy is that the onset threshold of the target device can be reduced on light illumination, allowing the synaptic plasticity learning and recognition acceleration at a lower power.Finally, under light exposure, the target synaptic device achieves random handwritten digit recognition with an impressive accuracy of 89.3% after only 500 training cycles, which means that the target has the ability to learn quickly.

| RESULTS AND DISCUSSION
The schematic diagram of a pre-synaptic neuron (axon) and a post-synaptic neuron (dendrite) is shown in Figure 1A.and a biological synapse is well-defined as a connection between the pre-and postneurons.As a synaptic device, the memristor with the structure of Au/[CNBmim]Cl/α-FAPbI 3 /SnO 2 /ITO (Figure S1) and synaptic simulation is also shown in Figure 1A, and the molecular structure of the [CNBmim]Cl is displayed in Figure S2.XRD patterns were used to carry out the crystal structure of the FAPbI 3 film.As shown in Figure S3B, the most intense peak detected at 13.9°c orresponds to (110) diffraction planes, and the peaks situated at 28.1 and 31.4°maybe attributed to (220) and (310) diffraction planes, which are typically indexed to α-FAPbI 3 . 26,27No change appears in the peak locations of the perovskite film with or without treatment with [CNBmim]Cl (target film or control film), indicating that [CNBmim]Cl does not significantly alter the crystal structure of α-FAPbI 3 , which is further confirmed by the UV-vis absorption spectra (Figure S9B and S9C).Moreover, the improved crystallinity of the α-FAPbI 3 film is demonstrated by the enhanced intention of diffraction peaks after being treated by [CNBmim]Cl.Furthermore, SEM of the control film is illustrated in Figure S4.As is shown in Figure S4A and Figure S4B, the grain size in the control film is small and inhomogeneous, and a large number of grain boundaries (GBs) are displayed in this film.At the GBs, the deep states' defects within the band gap could be generated by the dangling bonds from Pb 2+ or halide ions and act as charge recombination centers. 28Besides, defects are an important source of instability of perovskite devices, especially in synergy with the light, humidity, heat or oxygen, which accelerates the decomposition of perovskite. 29Therefore, we used the [CNBmim]Cl to treat the α-FAPbI 3 film to passivate the defects.The interaction between [CNBmim]Cl and α-FAPbI 3 , and the passivation mechanism is analyzed (Figure S5).As a result, compared with the control film, the grain size increased and the GBs reduced sharply obviously in the target film, further confirmed by the results of EIS and space-charge limited conduction (SCLC) (Figure S6) and Table S1 and the accompanying discussion in Supporting Information.
The RS switching characteristics of the control and target devices are investigated and shown in Figure 1B.As can be seen, both devices display evident bipolar resistive switching behavior.Only a few dozen stable I-V cycles can be maintained by the control device; however, the target device demonstrates a constant endurance of up to 550 cycles (Figure S7A) with enlargement of the storage window with an On/Off ratio (>10 2 ), and longer retention time of up to 20000 s, as shown in Figure 1C,D, which is comparable with or even better than that of the perovskite memristors have been reported (Table S2).And a distinct reduction in the maximum current in the negative reset operation (I RESET ) in the target device is observed, which is required for low-power consumption devices.In addition, we also tested the stability of the device at elevated temperatures (Figure S7B).Four stable LRS and HRS can be acquired by regulating the limiting current during the set process and the cut-off voltage during the reset process (Figure 1E).A compliant current of 10 mA is set to protect devices from a hard breakdown.As displayed in Figure 1E, when changing the I-V curve of the reset stop voltages (−1.0, −1.5, and −2.0 V) and setting the stop voltage of +2.0 V, the target device exhibits different I-V curves, indicating multi-level storage capacity, suggesting the potential application as an artificial synapse.The LRS and HRS at 0.2 V were further extracted separately and plotted (Figure 1F).The LRS is remarked as level 1, whereas the HRS in the reset process with applied voltages of −1, −1.5, and −2 V are remarked as levels 2, 3, and 4, respectively.The HRS between each level are clearly distinguished with ratios of 13 and 84, which is consistent with the results shown in Figure 1E, and the devices showed good endurance in different resistance states.
To further confirm the effect of [CNBmim]Cl treatment on α-FAPbI 3 film and analyze the current conduction mechanism, the LRS and HRS of the I-V curve were fitted (Figure 2A) according to the conducting mechanisms of Ohmic conduction and Schottky emission, respectively.For the fitting results of LRS in the target device, a well-linear relationship between ln(I) and ln(V) in the range from −0.05 to −0.8 V is obtained, and the corresponding slope of the curve is 1.02 (Figure 2B) using Equations ( 1) and (2), 30 indicating the LRS of the target device is based on an ohmic conduction mechanism. 31 In Equation ( 1) σ and μ represent the electrical conductivity and ionic mobility of the device, respectively; N C , E C , and E F are the effective density of states, conduction band, and Fermi level of the functional layer, respectively.In contrast, the dominant conduction mechanism of the device in the high resistance state is the Schottky emission mechanism (Figure 2C), and the variation of I-V curves with temperature was analyzed using Equation (3). 32 As shown in Figure 2D, according to the fitting curves at each temperature in the range of 300-340 K (Figure S8), the temperature-dependent I-V switching characteristic are all consistent with the Schottky emission mechanism. 33This result further confirms that the Schottky emission mechanism is satisfied in the HRS, and good thermal stability of the target device is also confirmed.Thus, with light illumination, the set voltage of the device can be controlled as low as 0.1 V, and the energy consumption of the memory device can also be reduced by photo-assisted switching.It is possible that the RS is a consequence of the migration and redistribution of I − , due to the activation energy of I − being much lower than that of Pb 2+ . 34During the SET process, I − ions are drifted to the top electrode of Au, leaving iodine vacancies (V I ˙) in the perovskite film to form CFs and eventually connecting the bottom and top electrodes under the electric field.During the RESET process, I − ions reverse migration and recombine with V I leading to the CFs rupture. 13,25It is reported that an iodine cluster in 3D α-FAPbI 3 is so stable that the CFs are difficult to fracture, 25 which maybe the reason for the high power consumption and unstable RS behavior in the control device.In our previous work, 35 passivation treatment on the perovskite film, the primary ion migration mechanism could gradually change to the Schottky emission mechanism, which is consistent with the RS conduction mechanism in the target device.
Theoretical research studies demonstrated that aggregation or dispersion of shallow defects can lead to an increase or decrease of the electronic conductivity of perovskite and manifest as RS behavior. 36The interface state is a key factor in the RS transition. 15Specifically, the initial voltage is 0 V as shown in Figure 3A.When a positive voltage (V SET ) is applied on the Au electrode (Figure 3B), the charges extracted from the bottom electrode can be trapped by the bulk defects in α-FAPbI 3 film until these defects are filled.Subsequently, a sharp increase in the current and the target device is set to the LRS.In contrast, with a negative voltage (V RESET ), the trapped electrons in the interface and bulk defects can be detrapped, and the target device is reset to the HRS. 14Practically, due to the passivation by [CNBmim]Cl, the defects in the bulk α-FAPbI 3 film and interface between the α-FAPbI 3 film and Au electrode are reduced obviously, which accelerates the separation of electrons and holes and confirmed by the photoluminescence (PL) spectra in Figure S9.Furthermore, the rapid charge separation promotes the rupture of CFs and the transition of HRS.
Biological neurons and specific neurotransmitters are stimulated not only by electricity but also by light.Here, the light stimulation with an AC electrical impulse signal is applied to the target device (Figure 4A), to obtain a high-order tunable synaptic plasticity, and further promote the memory and learning capabilities.It is obvious that the V SET in the I-V curve is shifted to the direction of low voltage under illumination (Figure 4B) and the forming voltage of the target device with illumination is lower than that in the dark (Figure 4C).Moreover, the power consumption of the target device is 45.17 pJ and 3.43 pJ under dark and illumination, respectively (Figure 4D,E) according to the calculating method in previous references. 37,38 illumination has a faster voltage-current response.When the target device is exposed to light, the photogenerated electric field is produced and the photogenerated charge carriers are induced in the α-FAPbI 3 film.Moreover, the photogenerated electric field is in the same direction as the external electric field with the V SET , 11,39 which is beneficial to the trapping and detrapping process and filament formation leading to the lower V SET and faster voltage-current response.The proposed RS mechanism is shown in Figure 4H,I.As the artificial synapses, the target device has the ability to promote synaptic activity at low power inspired by the aid of light stimulation. 11,40hen the electrical input signal is sent from the preneuron to the synapse, the synaptic weight changes and determines the output signal, causing an alteration in the dendrite structure of the post-neuron and a change in the ion transmission.The capacity to tune the synaptic weight is called synaptic plasticity, and it is known to be the fundamental function for the prompt adapting and learning process. 41Moreover, the long-term plasticity including long-term potentiation/depression (LTP/LTD) is dependent on the input pulses. 42LTP and LTD are two types of synaptic plasticity that are involved in learning and memory in the human brain.LTP strengthens the connection between neurons, while LTD weakens it.When these features are simulated using perovskite memristors, it is possible to use them in artificial neural networks to improve their performance in tasks such as recognizing handwritten digits.The effect of illumination on the LTP/LTD characteristics during the simulation of biological synapses in the target device was explored, and the device was subjected to electrical pulse conductance modulation under 70 consecutive positive (potentiation) and negative (suppression) conditions (Figure S10).The electrical conductivity of the device after illumination is enhanced as well as the modulable range.Moreover, a good linear fitting result is performed in Figure S10B.That is, the LTP and LTD curves of the target device under illumination exhibit a high degree of coincidence with the fitted curves compared with that in the dark condition.In addition, the nonlinearity of LTP and LTD under illumination is 5.41 S and 6.58 S, respectively, which are significantly lower than that under a dark condition (15.41 S and 24.89 S).This result indicates that the target device could simulate the LTP and LTD characteristics of synapses and the synaptic weight will be modulated when the device is under light conditions.
The low impedance of the device can be controlled by adjusting the compliance current, thus achieving a multilevel memory application.Figure 5A shows the I-V switching behaviors of the target device, where the stop voltage amplitude is boosted from −0.5 to −1.5 V during a negative bias sweep.And the Icc value increases from 1 to 10 mA during a positive voltage sweep, that is, the device has a gradual "RESET" and "SET" process.The target device exhibits good multistage storage performance, whereas the control device does not have that performance.The multi-level storage of the memristor was realized, which is the basic function to mimic the biological synaptic. 43Seventy consecutive boosting and inhibitory pulses were applied to the target device, with positive (negative) pulses representing the process of enhancing (reducing) synaptic weights corresponding to the changes in conductance (Figure 5B).Compared with the control device, an enhanced significance in the conductance with the larger variation range can be detected in the target device.The further analyzed of results in Figure 5B by linear fitting according to the following Equations (4-6). 44


where G LTP and G LTD are the conductance of the LTP and LTD processes, respectively; P is the pulse sequence set during the test and P max is the total number of pulses; and G max and G min are the maximum and minimum conductance of simulated synapses when switching between high and low conductance during long-term and short-term memory, respectively.These values were all obtained from our experimental data.In addition, A in the above formula is an important parameter of linearity and the value of A will affect the size of the "window" of the fitted image, and B is a function of A. Figure 5C,D is derived from the data in Figure 5B by applying Equations ( 4) and ( 5).Compared with the control device, the conductance of the target device has better linearity.And under the same pulse energy, the target device with good linearity is more likely to produce a change in conductance and simulate the function of synapses.
To realize the relevant neural functions in the synapse, the simulation of the synaptic plasticity of the device was investigated under pulsed electrical stimulation by paired-pulse facilitation (PPF) and enhancement of second excitatory postsynaptic potentials (EPSPs).PPF in neuroscience is a form of short-term plasticity manifested by EPSPs in rapidly evoked two closed spikes. 45,46And the PPF can be calculated by Equation (7). 47,48 During the tests, the amplitude of the two peaks was set to 1.0 V, the two read pulses were set to 100 ns with an amplitude of 0.2 V.The first and second applied currents are denoted by I 0 and I.The value of the EPSPs is closely related to the time interval between the two spikes applied to the pre-synaptic membrane.As the time interval increases, the enhancement effect caused by the second pulse becomes smaller, which is consistent with biological synapses.In addition, spike time-dependent plasticity (STDP) plays a key role in the cognitive behaviors of the nervous system and reflects the influence of pre-and post-synaptic stimulation on synaptic weights. 37STDP is also an important synaptic learning mechanism that regulates high-level neural activities and a research direction of memristor simulation synapse models. 49,50According to the Hebbian learning rule, 51 the time differences between the presynaptic spikes reaching the postsynaptic membrane (Δt post-pre ), that is, the timing and sequence of firing, affects the synaptic weights, which in turn regulates the strength of connections between neurons.If the previous synaptic spike is triggered before the next one, the synaptic weight increases and the strength of the connection between the two neurons increases.Conversely, if the weight decreases, then the connection becomes weaker.The synaptic weight change can be calculated by Equation ( 8). 52 where B and τ are the constant and time constant, respectively; ΔW 0 is the conductance value at the Δt limit; W 0 and W STDP are the synaptic weights before and after pulse application, respectively.To analyze the STDP of the target device, the pulse waveform shown in Figure S11 was chosen.As shown in Figure 5F, the numerical values of the synaptic weights vary with the time difference Δt post-pre .When the presynaptic membrane spike time arrives before the postsynaptic membrane (Δt post-pre > 0), the device is in SET, otherwise, the device is in RESET.All STDP data of passivated device fit well, and the target device as an artificial synapse successfully achieves the STDP properties of physiological synapses, which lay the foundation for multifunctional neuron-related computations.
For artificial synapses, the recognition of the image corresponds to a series of different current signals, which in turn causes the device to output a number with a different color.As shown in Figure 6A, an artificial optoelectronic synapse model is proposed, which is able to simulate the function of the human visual central system.Convolutional neural networks CNNs are one of the mainstream choices for memristive neuromorphic computing and edge computing.We applied the CNN to the target artificial synapses for random handwritten digit recognition (Figure 6B).The neural network consists of an input layer, X1, a convolutional layer, X2, and an output layer, X3. [53][54][55][56] Among them, X1 consists of an integrated array and the handwritten digit recognition process occurs at X2 and X3.To investigate the learning capability of the target synaptic device, the MNIST pattern recognition was simulated based on the fitting results of the LTP and LTD in Figure 5 and calculated by convolution.It is possible to obtain the digital recognition accuracy, loss rate, and the corresponding images after identification under illumination.As shown in Figure 6C,D, under illumination, the recognition accuracy of the target device also significantly improves compared to the dark state.When the learning training cycle increases to 500 times, the recognition accuracy of the target device for handwritten digits increases linearly, while the recognition loss rate also decreases accordingly.Moreover, the recognition accuracy of the target device is significantly higher than that of the control device, reaching 89.3%.In the recognition process, training has an important effect on the loss rate and accuracy rate (the training process is shown in Figure S12).The impact of training on the image recognition ability of this device may be appreciated from Figure 6E-J.Figure 6E shows a randomly selected handwritten digit image "6" from the MNIST database.
Before training, the image after the recognition of the target device was blurred (Figure 6F), and cannot obtain a clear number "6."It is not possible to recognize any of the 10 numbers from 0 to 9, and accurate recognition cannot be achieved.This can be determined by identifying outcome-type probability histogram proofs (Figure 6G) and the synaptic weights are randomly distributed before the device is trained.However, the randomly selected number "5" (Figure 6H) can be accurately identified (Figure 6I), which is also apparent from the number probability histogram (Figure 6J) recognized from the target device after 500 training cycles.Among the ten numbers from 0 to 9, the number "5" has the highest probability of being recognized, that is, the synaptic weights are concentrated, which realizes the simulation of part of the function of the human visual system.

| CONCLUSION
In summary, the 3D trigonal FAPbI 3 (α-FAPbI 3 ) with excellent photoactivity and thermal stability was applied as the active layer for the artificial synaptic electronic device with Au/[CNBmim]Cl/α-FAPbI 3 / SnO 2 /ITO structure.The interface between the α-FAPbI 3 layer and Au electrode was modified by [CNBmim]Cl, which reduced the defects in GBs of perovskite and the charge recombination center.This enables the device to accelerate the detrapping of electrons and promote the CFs capture during the RESET process, which leads to the fast transition from LRS to HRS and stable RS behaviors with an On/Off ratio (>10 2 ).In addition, the retention time and endurance of this device are up to 10000 s and 550 cycles, respectively.In contrast, the storage window of the untreated α-FAPbI 3 based device is only a dozen, and can only run a few I-V cycles before failure.On the other hand, due to the passivation of interface defects in α-FAPbI 3 by [CNBmim]Cl, the sensitivity of the perovskite layer to the environment is reduced, which inhibits the decomposition of perovskite caused by external water and oxygen and improving the stability of perovskite devices.Moreover, with the aid of light, the device has low energy consumption (~3.43 pJ), which is lower than that of many hybrid methylammonium lead halide perovskites-based memristors.And this device is capable of mimicking basic synaptic behaviors including LTP, LTD, and STDP.Finally, this device is applied to neuromorphic calculation; with the light-assisted revision of the synaptic weights, the recognition accuracy of random handwritten digits is up to 89.3%, which is infrequently under low training times (500 training cycles).This work introduced a simple interface engineering method to realize an artificial synaptic electronic device based on α-FAPbI 3 and realize the transition from LRS to HRS of the memristor, which is the first report of an α-FAPbI 3base artificial synaptic electronic device, to the best of our knowledge.This work provides a new way to the material types of OHPs-based memristors or artificial synaptic devices besides δ-FAPbI 3 and hybrid methylammonium lead halide perovskites and expands the application prospect in the field of neuromorphic computing.

F
I G U R E 1 (A) Schematic of a synapse and the structure of the target device, (B) I-V curves for the control and target device.(C) Retention testing of the target devices, (D) Resistance distribution in the HRS and LRS for both devices, (E) I-V curves of the target device with different stop voltages, (F) The measured multilevel storage characteristics for target devices.HRS, high resistance state; LRS, low resistance state.
Fitting plots of target device: (A) I-V curve, (B) low-resistance state LRS, (C) high-resistance state HRS, and (D) temperature dependence curves of the RS mode in the temperature range of 300-340 K. HRS, high resistance state; LRS, low resistance state; RS, resistive switching.
Figure 4F,G shows the degree of delay between the input and output of the target device.The delay time (Δt) decreases after illumination, indicating that the target device under F I G U R E 3 Schematic of the tentative RS mechanism of the (A) initial state, (B) SET (write), and (C) reset (erase) processes of the target device under dark.

F
I G U R E 4 Photo-response of the target device and effect of illumination on the simulated neural synapse.(A) Schematic showing the measurement of the photocurrent of the target device.(B) Under light and dark conditions: I-V plots of the target device measured under dark and illumination conditions under the voltage mode, (C) Electroforming process of the device.Under the pulsed conditions: the power consumption of the target device under (D) dark and (E) illumination conditions.Typical response curves of the target device to set pulses under (F) dark and (G) illumination conditions.(H) and (I) Proposed RS mechanism.RS, resistive switching.

F
I G U R E 5 (A) I-V curves of the target device under a continuous cut-off voltage of −0.5 to −1.5 V and continuous limited current of 1 to 10 mA.(B) Illumination under the same pulse conditions (potentiation and depression) of the control device and target device, the change curve of the synaptic weight during the simulation of biological synapses.(C) Fitting diagram of the actual measured conductivity and simulated synaptic weight of the control device and (D) the target device.(E) Simulated images of the PPF effect were obtained by applying forward pulses (1 V, 100 ns) to the resistor with different time intervals of the target device.(F) Calculation of the Hebbian STDP map in the target device based on the experimental data (all measurements were under illumination).

F
I G U R E 6 (A) Schematic showing a handwritten digit observed through the human visual system.(B) Schematic of an artificial neural network.Graphs of recognition loss rate (C) and accuracy rate (D) of the control device and target device under light and dark states.(E) Input and (F) output images of the target device before training.(G) The probability of outputting the result after identifying the digits (0-9) without training.(H) Input and (I) output images after device training.(J) The probability of outputting the result after identifying the digits (0-9) with training.