A Dynamic Memory for Reservoir Computing Utilizing Ion Migration in CuInP2S6

Time‐series analysis and forecasting play a vital role in the fields of economics and engineering. Neuromorphic computing, particularly recurrent neural networks (RNNs), has emerged as an effective approach to address these tasks. Reservoir computing (RC), a type of RNN, offers a powerful and efficient solution for handling nonlinear information in high‐dimensional spaces and addressing temporal tasks. CuInP2S6 (CIPS), a van der Waals material with ion conductivity, shows promise for sequential task processing. Here, a synapse device based on CIPS is demonstrated that exhibits temporal dynamics under electrical stimulation. By controlling Cu+ ion migration, this study successfully emulates synaptic performance, including potentiation and depression characteristics, and RC. Migration of Cu+ ions is confirmed using piezoresponse and Kelvin probe force microscopy. The device achieves low normalized root mean square errors (NRMSE) of 0.04762 and 0.01402 for the Hénon map and Mackey‐Glass series tasks, respectively. For real‐life time‐series prediction based on the Jena temperature database, an overall NRMSE of 0.03339 is achieved. These results highlight the potential of CIPS ion conductivity for real‐time signal processing in machine learning, expanding applications in neuromorphic computing.

In this work, we investigated the dynamic responses of CuInP 2 S 6 synapses utilizing a two-terminal memristor, which exhibited short-term memory for high-performance RC appli-cations.The synapse displayed both volatile performance and a time-dependent response.Our investigation focused on the natural ferroelectricity of CuInP 2 S 6 and the transport characteristics of the memristor.Furthermore, we employed a combination of Piezoresponse Force Microscopy (PFM) and Kelvin Probe Force Microscopy (KPFM) techniques to verify the migration of Cu + ions and confirm the underlying mechanism.We successfully simulated synaptic plasticity, including long-term potentiation (LTP), long-term depression (LTD), and double-pulse facilitation (PPF), in the copper ion motility-tunable electron transport of CIPS.To assess the ability of the ion migration-RC system, we conducted pattern recognition and waveform classification tasks, which demonstrated high performance in both training and testing.Additionally, the CIPS device exhibited excellent performance in time-series prediction, chaotic systems analysis, and real-time neural activity analysis.By leveraging the dynamic memristor's competence, we achieved high accuracy in tasks such as the Hénon map and Mackey-Glass series, with a lower normalized root mean square error (NRMSE) of 0.04762 and a prediction error (NRMSE) of 0.01402, respectively.Furthermore, the time-series prediction of real-life data based on the Jena temperature database resulted in a prediction error (NRMSE) of 0.03339.

Materials Characterization
12a,17b] Figure 1b shows the electrical measurement process, and voltage was applied to the top and bottom electrodes.The CIPS nanoflake was exfoliated from the bulk and transferred to the conductive Ni electrode (details information see Experimental Section).The Raman spectrum is characterized by the Renishaw inVia Raman with a 532 nm wavelength laser to confirm CuInP 2 S 6 .2c,4,13c,31] The exfoliated CIPS single crystal sample quality is characterized by SEM-EDS (Scanning electron microscopy (SEM) and energy-dispersive X-ray spectroscopy (EDS)) mapping as shown in Figure S1 (Supporting Information), which shows that each element is uniform and located on the flake.
Then, we further discuss the pristine ferroelectricity performance by using PFM. [2,13]Figure 1d represents the illustration of the strategy for PFM, in which a voltage is applied to the conductive tip on the sample.The voltage establishes an extended electronic field between the tips and substrate.The ferroelectric material shows a response under this field, which will expand or shrink depending on the electric field direction.Recent studies reported that CIPS polarization response was attributed to the Cu + migration. [2,12,18,32]PFM was used to manipulate and detect the ferroelectric polarization of CIPS flakes on the Nickle substrate under different thicknesses (Figure S6, Supporting Information).Figure 1j shows the typical hysteresis loops of phase and amplitude versus applied voltage where the strong phase changes up to 180°for a 160 nm CIPS flake and the amplitude curve displays a butterfly-like shape.To comprehensively assess the ferroelectric domain of 189 nm CIPS flake, a sequential procedure is followed.Initially, a −12 V voltage is applied, resulting in a downward orientation of the domain.Subsequently, a +12 V is applied to the middle area, inducing an upward domain orientation.To revert the domain back to a downward state, another −12 V voltage is employed.Finally, a small voltage is applied to enable the reading of polarization across the entire area.Figure 1f,g depicts the amplitude and phase images, respectively.Notably, the oneby-one pattern remains remarkably clear even after one week, as evidenced by Figure S7 (Supporting Information).This observation provides compelling evidence that the ferroelectric polarization effectively undergoes flipping in response to an external electric field.Additionally, to verify the ferroelectric properties, we employed the Positive-Up-Negative-Down (PUND) and Dynamic Hysteresis Measurements (DHM) methods, as illustrated in Figure S5 (Supporting Information).

Electrical Characterization and Mechanism of Mobile Cu + Ions
Utilizing the ferroelectric and semiconducting characteristics of CIPS, we fabricated a metal-semiconductor-metal (MSM) structure memristor, as depicted in Figure 1b.To investigate its behavior, a range of sweep voltages from ±1 to ±3 was applied to the MSM device, with both contacts composed of Ni.The resulting I-V (current-voltage) characteristics are presented in Figure 2a and Note S1 (Supporting Information, showing the I-V result of device-to-device variation).Initially, when the sweep range is ±1 V, we observe an almost symmetric I-V behavior.However, as the sweep range increases, a distinct asymmetric I-V curve emerges, indicating a transition from threshold resistive switching to self-rectification. [33]The switching ratio is shown in Figure S2 (Supporting Information) and the I HRS /I LRS is ≈400 at a reading voltage of 1.5 V. Considering a sweeping range of ±2 V, when a positive sweep is applied from 0 to +2 V, the device undergoes transitions from a high resistance state (HRS) to a low resistance state (LRS) at a threshold voltage (V th ) of ≈2.0 V.During the back sweep, the device returns to HRS again, which indicates that the hysteresis loop is originated from the bidirectional threshold resistive switching effect similar to that observed in SnS [5] and previous CIPS work. [2,18]Such RS behaviors have been exploited for selectors, [34] logic applications, [35] and neuromorphic computing. [16,36]The performance has indicated that the CIPS device has the advantage for multiple applications, in addition, our device shows an operation speed rate of 200 μs (see Note S1, Supporting Information for the switching speed).
Next, we focused on investigating the mechanism in vertical MSM (Ni/CIPS/Ni) devices.Figure 2b and Figure S4 (Supporting Information) present the cycle-to-cycle retention measurements conducted at a voltage of ±3 V and read at +1.5 V. Additionally, Figure 2c illustrates the cumulative probability of the conductance values for both the high resistance state (HRS) and low resistance state (LRS) obtained from Figure 2b.After 90 cycles, the on/off ratio was observed to be ≈85.To elucidate the underlying mechanism of the device, a pulse stream was applied to the MSM structure.Specifically, a train of 20 pulses, each with a pulse width and interval of 10 ms, and an amplitude of 3.0 V relative to a base voltage of 0 V, was employed.The device exhibited an increasing response current with each subsequent pulse.However, once the voltage dropped back to the base voltage, the current decreased to a significantly lower value, as illustrated in Figure 2d.These phenomena demonstrate a time-dependent and volatile behavior in this scenario.A closer examination of the response, depicted in Figure 2e, reveals a zoomed-in image with a time window ranging from 0.037 to 0.052 s.Notably, two distinct stages are observed within this time frame.In stage I, there is a rapid increase characterized by a steep, cliff-like slope.Subsequently, in stage II, the growth is slower compared to stage I.2b] This two-stage behavior is closely linked to the migration of Cu + ions along two distinct paths.In stage I, Cu + ions predominantly migrate in-plane (IP), while in stage II, the Cu + ions migrate outof-plane (OOP).12b,37] The disparity between the two migration paths can be attributed to the variation in the activation energy required for hopping or motion.2b,17a,18a] This discrepancy in activation energy accounts for the observed difference in behavior between the two paths.The lower barrier associated with the in-plane (IP) migration path indicates that it requires less energy to drive the movement of Cu + ions.Consequently, in stage I, the current experiences a rapid boost, as evidenced by the fast increase in current attributed to IP Cu + ion migration.On the other hand, in stage II, the migration involves a longer distance and higher energy requirement, resulting in a slower growth rate.This accounts for the observed slower growth during stage II, as it takes more time for the Cu + ions to traverse the longer distance and overcome the higher energy barrier.Hence, the I-V behaviors can be described as follows, as illustrated in Figure 2f.Initially, during the sweep range from 0 to +2 V and back to 0 V, the Cu + ions migrate from the anode to the cathode and accumulate near the cathode surface.This accumulation leads to an enrichment of Cu + ions near the cathode while causing a deficiency of Cu + ions near the anode.Consequently, during the positive sweep, the device undergoes a switch from the initial high resistance state (HRS) to the low resistance state (LRS) as indicated by (i) to (ii).Applying an opposite voltage creates an electric field in the opposite direction, causing the Cu + ions to migrate back, leading to a switch from (ii) to (iii), transitioning from LRS to HRS.By continuously applying an electric field, the device can transition back to LRS, as observed during the transition from (iii) to (iv).Furthermore, previous studies have reported that the application of a larger voltage drives the mobile Cu + ions toward the cathode, leading to the formation of an asymmetric interfacial barrier.2d,38] Consequently, under the influence of a high electric field, the Cu + ions are capable of hopping, which induces phase separation within the material.
Based on the aforementioned discussion, the observed I-V behaviors can potentially be attributed to Cu+ migration.In order to further confirm this mechanism, we employed a combination of PFM and KPFM.PFM is a well-established technique for characterizing ferroelectric materials, as mentioned earlier.On the other hand, KPFM is a powerful method used to measure various electrical properties of materials, including surface potential and work function.Next, we applied PFM and KPFM to further investigate the Cu + migration under the electrical field.The surface potential [2b] difference can be defined as: where W tips and W CIPS are the work functions of conductive tips and CIPS, respectively, and e is the electronic charge.Hence, the work function difference could be defined as: Figure S9 (Supporting Information) displays the CIPS topography, potential, and work function before any treatment.Following the application of PFM treatment (Figure S8, Supporting Information), the domain becomes highly visible, as depicted in Figure S8a,b (Supporting Information), achieved using a writer voltage of −12 V. Interestingly, after the writing process, the surface potential within the writer area is found to be 40 mV higher than the untreated regions, as shown in Figure S8e,f (Supporting Information).This observation provides evidence that the movement of Cu + ions occurs, leading to an increase in the potential of the treated region.In light of the preceding discussion, it is apparent that the conductive behavior and switching of Cu + ions play a fundamental role in the resistive switching (RS) behavior, thereby serving as a fundamental property of ion current.

Temporal Dynamic of CIPS Memory
Given the unique potential properties of CIPS, particularly its temperature response and volatile nature, it possesses the capability for RC.This is attributed to two essential characteristics of the RC system: non-linear response and short-term memory (STM). [39]STM and long-term memory (LTM) are fundamental components of the biological memory system. [40]STM is characterized by its brief retention period, usually lasting only a few seconds to minutes.In contrast, LTM exhibits a significantly extended capacity, allowing it to retain memories for periods exceeding a day and potentially extending far beyond.Figure 3a illustrates the physical reservoir, wherein the temporal input is transmitted through neurons with random and nonlinear connections, and the input can be read out using straightforward algorithms.The crucial aspect of the RC system lies in the decay time of STM, which can be measured by observing the postsynaptic response of PPF.In Figure 3b, it is evident that the decay time, measured at V Base = 0 V, is ≈43.6 ms, primarily resulting from the fading effect.However, an intriguing observation is the transition from STM to LTM as the base voltage increases to 1.5 V, as seen in previous experiments.This transition may be attributed to the electrical field, which provides a sustaining voltage for Cu + ions.Consequently, the decay time significantly increases to ≈125.3 ms with a basement voltage of 1.5 V, as depicted in Figure 3c.We then delve into the synaptic behaviors for mimicking artificial synapses.LTD and LTP are two instances of synaptic plasticity, [41] which have been observed in our device as well.They play a crucial role in learning and memory by modifying neural activity.LTP is an enhancement behavior that occurs through stimulation.LTD is characterized by a persistent decrease in synaptic plasticity under opposite stimulation.The dynamic response of the device was investigated by applying 20 positive pulses (+3 and +2.5 V) and 20 negative pulses (−3 and −2.5 V) with varying basement voltages (ranging from 0.5 to 1.5 V in increments of 0.5 V) and pulse parameters of 10 ms pulse width and 5 ms pulse interval.Figure S3 (Supporting Information) showcases the current response, defined as postsynaptic current (PSC), which increases with the pulse stream.Furthermore, the dynamic behavior of the memristor in response to electric pulses is depicted in Figure 3d.The input pulse amplitude ranges from 0.5 to 3.0 V in increments of 0.5 V, and a pulse width of 10 ms is applied to induce dynamic states.
The spatiotemporal signal processing characteristic of RC computing is demonstrated by the multiple reservoir states depicted in Figure 3e,f.In these figures, input signals were applied, and dynamic modulation was achieved with the following inputs: (0111), ( 0110), (1111), and (0011).The device's response is a result of the combined effects of electrical stimulation and decaying memory.Starting from similar current values, the four inputs, each consisting of four electrical pulses, yield four distinct final current amplitudes.The migration of Cu in the presence of "1" generates a corresponding current, leading to an increase in current, while the signal of "0" tends to restore the conductance.Thus, for a given electrical input, the complex state of the memristor can be determined by the combination of "0" and "1".A total of 16 different input signals can be generated, with an example of a "0" input being (0111) as shown in Figure 3g.After writing, a read pulse of 3 V is added.Due to the fading memory effect, historical information is encoded in the output current, allowing for information mapping with input patterns.Figure 3h illustrates the read current reflected by the memristor after passing through a [5 × 4] pattern corresponding to the digit "9".The response current depends on the input stream, indicating different reservoir states.The individual reservoir states for "0" to "9" are shown separately in Figure S10 (Supporting Information).To further examine the dynamic response of our device, we constructed 50 patterns to elicit responses for different digits.All 50 patterns (30 for training and 20 for testing) are presented in Figure S11 (Supporting Information), where the current is fed into the reservoir for training and testing.We utilized the Python environment along with simple logistic regression [22,26,27] to determine the training × 5 input pattern representing digit "9" with binarized '0′ and '1′ bits correspond to gray and black pixels, respectively.The response currents sensed by memory after each pulse stream are fed to reservoir for training and testing.The output layer contains 10 digits labeled from 0 to 9. h) Read currents of 5 patterns corresponding to digit '9′.i) Variation of accuracy after iterations of the test process.j) Interfered confusion matrices between desired output digits and predicted train and test phase output.
weights.The training dataset was processed as an input vector by the readout software, with the output layer representing the corresponding numerical labels.Remarkably, we achieved a training accuracy of 100% within five iterations, indicating that our CIPS-RC system is capable of effectively processing all 30 train-ing patterns.Once the weight values were obtained, the remaining 20 patterns were sent to the reservoir for model testing.To assess the system's robustness, we introduced some noise and removed certain pixels in the test samples.As a result, the pattern recognition task achieved a maximum accuracy of 80% after five iterations (Figure 3i,j).The number of correctly predicted digits is depicted in the matrices in Figure 3f.Out of the 20 test datasets, four images were incorrectly predicted, and further details can be found in Figure S11 and Note S2 (Supporting Information).

Time-Series and Chaos System Prediction
To further validate the spatiotemporal information processing capabilities of CIPS-RC, we employed a cyclic system to handle time series tasks.The structure of the cyclic reservoir system is depicted in Figure 4a. [20,26,27]In this structure, the input signal undergoes a pre-processing step through a time-multiplexing process.This process involves introducing a delay time loop () to generate multiple reservoir states/nodes, which are separated by intervals of delay time, as depicted in Figure 4a.The reservoir nodes can be trained and tested using a very simple linear regression approach. [20,27]We conducted waveform classification to verify the performance of the temporal task.Detailed information about the multiplex process can be found in Figure S12a,b (Supporting Information).The waveform classification results, shown in Figure 4b, were obtained using M = 10 with five parallel reservoirs and a time delay  of 10 ms.The results exhibit a high coefficient of determination (R 2 ) accuracy of 0.9080 and a low normalized root mean square error (NRMSE) of 0.2688.The results for M = 5, 10, 20, and 40 can be found in Figure S12d,e (Supporting Information).As the number of parallel reservoirs increases, the NRMSE decreases, indicating improved accuracy.Additionally, the increased number of parallel reservoirs provides more virtual nodes, which can enhance the system's endurance and contribute to higher accuracy.
The time-series task is known to be more challenging due to the cumulative effect of prediction errors, which can lead to significant divergence from the ground truth.To evaluate the performance of the CIPS-RC system in time-series forecasting, we employed benchmark tasks such as the chaos system.The chaos system is particularly suitable for testing time-series prediction capabilities as even small errors can cause divergence in the predictions.Additionally, we conducted experiments using the Hénon map and Mackey-Glass task as examples to assess the system's ability for long-term prediction in time series.These tasks allowed us to evaluate the system's performance in capturing complex dynamics and making accurate predictions over extended time horizons.Hénon map [20,42] can be described by the following: (3) where a = 1.4 and b = 0.3.By solving the above equations, we can get the input x(n), and the target output is x(n+1).Using these equations, we obtain a Hénon map dataset with a length of 1000, which is being used for training and testing.The preprogramming mask employed in this study shares similarities with the waveform classification approach discussed earlier.The pre-processed signal is multiplied by a mask with a length of M, resulting in an input stream with a time delay of 10 ms ().The training and testing procedures involve utilizing a linear regression model and determining appropriate parameter values.Notably, our RC system exhibits exceptional performance in terms of time-series prediction accuracy.Figure 4c presents an illustrative example using M = 10 and Vmax = 3.0 V.The test results (prediction) are compared to the target values for the first 200time steps, exhibiting a low NRMSE of 0.04762 and a high R 2 value of 0.9997.This achievement is attributed to the CIPS-RC reservoirs with four paralleled reservoirs, aligning with previous findings. [20,42]Additionally, Figure 4d displays a 2D representation of the Hénon map, demonstrating a significant overlap.The impact of the number of paralleled reservoirs is also investigated (Figure 4e), revealing a decrease in NRMSE from 0.1 to 0.047 as the number of paralleled reservoirs increases from one to four.For a long-term prediction (1000-time steps) with a mask length of 10 and only one reservoir (Figure S13a, Supporting Information), the NRMSE is ≈0.2015.Moreover, device-to-device variation is examined (Figure S13c, Supporting Information), and detailed information on M = 5, 10, and 20 can be found in Figure S13b (Supporting Information).
In addition to the findings, we also explore the capability of the CIPS-RC system in handling chaotic systems.25b,30b,43] This time series is defined by the following equation: The Mackey-Glass time series serves as a comprehensive benchmark for testing forecasting tasks, given its deterministic nature and inherent difficulty in prediction.In our study, we choose specific parameter values, namely  = 0.2,  = 0.1, and n = 10.An important factor in this chaotic system is the time delay parameter, .For values of  >17, the system exhibits chaotic behavior.To investigate the system's performance, we select two different time delay values, namely  = 15 and  = 30, as described in the methodology section. [43]In order to enhance the prediction accuracy, we employ two techniques to expand the reservoir.First, we design masks of varying lengths to add virtual nodes, thereby improving the accuracy of the predictions.This is demonstrated in Figures S14 and S15 (Supporting Information) for both  = 15 and  = 30 S, we utilize paralleled reservoirs based on the response variable of different masks, similar to the approaches used in the classification and Hénon map tasks.The results of different mask lengths are shown in Figures S14b and  S15b (Supporting Information), and notably, they exhibit excellent prediction accuracy, with an NRMSE of ≈ 0.027 with M = 10.The performance of different devices can vary, as shown in Figure S15c (Supporting Information) for the device-to-device variation under M = 10.Furthermore, the influence of paralleled reservoirs on the Chaotic system is discussed in Figure 4f-h (see Note S3, Supporting Information for the prediction compared with other RC systems).With four paralleled reservoirs, the R 2 value approaches the ideal value of 0.9979 for both the training and testing processes, indicating excellent results.Additionally, a small NRMSE of 0.014 has been observed in this scenario.Overall, our CIPS-RC system demonstrates outstanding performance in the Chaotic system, showcasing its strong capabilities in parallel processing (see Note S4, Supporting Information, for the energy consumption with other RC systems).

Real-Life Time-Series Prediction of Daily Temperature
To assess the real-life signal processing capability of the CIPS-RC system, we conducted a benchmark using the Jena temperature database, which provides daily temperature data as the input signal. [27]The original Jena temperature database contains recorded temperature values for each time period, with data sampled every 10 min (see the Experimental Section for more details).For our experiment, we applied the same preprocessing procedure as before, utilizing a time-multiplexing process with a mask (m = 10) composed of −1 and 1 elements.Within each time interval ( = 10 ms), the input sequence was transformed into ten virtual nodes, resulting in a dynamic response from our RC system.The corresponding current generated by the applied signal was recorded, resulting in ten conductance states.The preprocessed input values, the values after the mask process, and the recorded responses are depicted in Figure S16 (Supporting Information).The corresponding current is then fed into the reservoir, and linear regression is employed for the training and testing process (see Experimental Section for further details).We specifically selected a period of 2000 h from the database for training and testing, encompassing the timeframe from 23.05.2009 (16:00:01) to  14.08.2009(23:00:01).The first 1000 h was allocated as the training dataset, where the input layer was applied and the weights and biases were determined.Subsequently, the remaining 1000 h were utilized for testing our model.Remarkably, our RC system achieved significantly lower prediction errors for both the training and test datasets, with values of 0.04057 and 0.03339, respectively.Additionally, the R 2 coefficients exhibited high accuracy, measuring 0.9825 for training and 0.9776 for testing.
For the purpose of long-term prediction, we extended our analysis to a considerably longer timeframe of 12000 h.This duration was divided into six distinct periods, each spanning 2000 h (further details can be found in the Experimental Section).The initial period, referred to as "summer," was used for training the system using a single reservoir configuration (M = 10, Vmax = 3.0 V).Specifically, this training phase focused on capturing the temperature variations during the summer season in Jena, which ranged from 4 to 27 °C.Subsequently, the following five seasons, namely autumn, winter, spring, and another summer, were designated for testing the performance of our model.The results obtained for these four seasons are depicted in Figure 5b-e, and a more comprehensive representation of the predicted and desired output for the "autumn" period is presented in Figure S17 (Supporting Information).We observed high accuracy and low NRMSE values in three seasons: autumn, spring, and summer.The R 2 values for these seasons were 0.9753, 0.9711, and 0.9766, respectively, all exceeding 97%. the NRMSE values were 0.06097, 0.07412, and 0.03327, indicating a good level of prediction accuracy.A summary of the R 2 and NRMSE values for all five seasons is presented in Figure 5f.However, during the winter period II, we observed higher NRMSE values (≈0.9778) and lower R 2 values (≈0.8840) compared to the other seasons.This discrepancy can be attributed to the temperature range falling outside the scope of the training database.Notably, we observed a significant improvement in accuracy during period IV compared to the other seasons.Additionally, the predictions during the beginning of autumn and the end of spring exhibited a better fit than other parts, possibly due to better alignment with the temperature ranges covered in the training dataset.

Conclusion
In conclusion, we have successfully fabricated the CIPS memristor and demonstrated its effectiveness in realizing a high-performance RC system through Cu + ion migration.The combination of PFM and KPFM techniques has provided direct evidence of the surface potential differences, confirming the involvement of Cu + ions in the observed I-V properties.The dynamic response of our device relies on ion conductivity, and the CIPS memristor exhibits both LTP and LTD, thus expanding the potential applications of artificial synapses and RC.Through the design and implementation of a mask process, our reservoir system effectively processes temporal signals, as evidenced by its performance in the Hénon map and Mackey-Glass time series prediction task.Furthermore, our system demonstrates high accuracy in real-life time temperature prediction, achieving a low prediction error (NRMSE) of 0.03339 using the Jena temperature database.Overall, the results obtained from the CIPS dynamic memristor highlight its potential as a robust platform for reservoir computing applications in time series analysis.Moreover, these findings open avenues for exploring its utilization in neuromorphic computing, presenting exciting prospects for future research and development in this field.

Experimental Section
Material Characterization: The single crystal flakes were analyzed using Renishaw inVia Raman Scattering Spectroscopy to confirm the Raman spectra.The measurements were conducted at room temperature in ambient air, using a 532 nm laser.Elemental composition was characterized using a field-emission scanning electron microscope (SEM) (Hitachi Regulus 8230) operating at an acceleration voltage of 5 kV.The thickness of CIPS flakes was determined by atomic force microscopy (AFM) using the Park System NX20 and Bruker Dimension Icon instruments.Phase and voltage loop measurements were performed using the Park AFM System NX20, which included a built-in PFM function, and local writing was done using the Bruker Dimension Icon system.KPFM measurements were conducted on the Park AFM System NX20.
Device Fabrication: The bottom and top electrodes were defined using a standard photolithography process with electron-beam lithography (EBL) using the Raith EBPG-5200 system.A 20 nm thick Ni layer was deposited on a p++ Si/SiO 2 (285 nm) substrate using the AJA UHV system E-beam Evaporator to create the bottom electrode after EBL.The same process was used to fabricate the top electrode, with a 40 nm Ni layer.The transfer of the materials was performed using the 2D Material Transfer Station (HQ graphene).
Database: The Hénon map dataset was generated by solving the corresponding function to obtain the X and Y values.The Mackey-Glass time series dataset for the Chaos System was obtained using the Runge-Kutta method for numerical integration.The Jena Climate database, recorded by the Max Planck Institute for Biogeochemistry in Jena, Germany, provides data at 10 min intervals.This dataset included ≈14 variables, such as temperature, pressure, and humidity, recorded from January 10, 2009, to December 31, 2016.For training and testing purposes, the dataset was resampled at hourly intervals from the original data.Figure 5a represents the average temperature recorded over a period of 2000 h, spanning from May 23, 2009 (16:00:01) to August 14, 2009 (23:00:01).For the longterm prediction task, the setting dataset from May 23, 2009 (16:00:01), to  November 30, 2010 (23:20:01) was divided into six, three-month periods, corresponding to the sequential order of summer-autumn-winter-springsummer-autumn.
Electrical Characterization and Parameter Extraction: Electrical measurements were conducted using the Keysight B1500A semiconductor analyzer under dark ambient conditions at room temperature.This included DC I-V and AC I-V measurements.Voltage pulse streams were generated using the built-in waveform generator fast measurement unit (WGFMU) in the Keysight B1500A analyzer.The polarization response hysteresis loop was measured using the aixACCT system TF ANALYZER 3000 i. Pattern recognition in Figure 3g was achieved using logistic regression.For time series forecasting tasks, linear regression algorithms were employed for RC training on daily temperature (Figure 5) and waveform classification (Figure 3h), as well as for the Hénon map and Mackey-Glass time series system.The Scikit-learn package in Python was utilized for regression training.

Figure 1 .
Figure 1.a) CIPS crystal structure.b) Schematic of the measurement method for a single device in the crossbar array.c) Raman spectrum of exfoliated CIPS flakes on SiO 2 substrate.d) PFM measurement setup with conductive tip on an electrostatic cantilever.e) Equivalent hysteresis loops of phase and amplitude.f-g) OP-PFM phase image of a 189 nm CIPS flake after writing the pattern by reverse DC bias (−12 V/+12 V).The OP-PFM amplitude image is shown.

Figure 2 .
Figure 2. a) I-V curves measured with varying sweep range, Vmax ranges from 1.0 to 3.0 V. b) The I on and I off read at V D = 1.5 V and switching ratio versus cycle number.c) Cumulative probability of both HRS and LRS conductance values from b). d) Current measurements under a multiple-bias cycle.e) A zoomed-in view of the green-shaded area in d).The schematics in the inset show the Cu + migration paths in the OOP and IP directions.f) Schematics of the dynamic ion migration during RS switching.The i, ii, iii, and iv are four states in the I−V curve of a).

Figure 3 .
Figure 3. a) Diagram of physical reservoir architecture.b) Schematic illustration of the current decay under different basement voltage (V Base ).c) The diagram shows the extracted  at various V Base .The current decay with time follows a simple exponential relationship and the characteristic time  obtained by fitting is 43.58 ms with the V Base = 0 V. d) Modulation of the current states with the pulse trains with pulse width of 10 ms and pulse interval of 5 ms.e) Response of a CIPS memory with a different pulse stream.f) Response to sixteen different states.g) A 4× 5 input pattern representing digit "9" with binarized '0′ and '1′ bits correspond to gray and black pixels, respectively.The response currents sensed by memory after each pulse stream are fed to reservoir for training and testing.The output layer contains 10 digits labeled from 0 to 9. h) Read currents of 5 patterns corresponding to digit '9′.i) Variation of accuracy after iterations of the test process.j) Interfered confusion matrices between desired output digits and predicted train and test phase output.

Figure 4 .
Figure 4. a) Diagram of cyclic reservoir architecture.b) Waveform classification with effective cyclic reservoir systems.c) Hénon map prediction results obtained by the memristor-based cyclic RC system, where the black line represents the ideal target and the red line represents the experimental output from the RC system.Test parameters are set to be M = 10, Vmax = 3.0 V. d) 2-D display of the predicted results under different test parameters.e) Hénon map showing accuracy with 4 paralleled reservoirs (m = 4).f) Coefficient of determination of Mackey-Glass time series prediction.g) Mackey-Glass time series forecasting results, 2-D display, and error rate.h) Mackey-Glass time series showing accuracy with 4 paralleled reservoirs (m = 4).

Figure 5 .
Figure 5. a) Desired and predicted data of daily temperature recorded in a period of 2,000 h experimentally obtained by CIPS memory-based RC system.b-e) A long-term prediction with different periods corresponding to autumn, winter, spring, and summer.f) Summarized R 2 and NRMSE of the testing process with 2000-time steps in each period after training the first period.