Reconfigurable halide perovskite nanocrystal memristors for neuromorphic computing

Many in-memory computing frameworks demand electronic devices with specific switching characteristics to achieve the desired level of computational complexity. Existing memristive devices cannot be reconfigured to meet the diverse volatile and non-volatile switching requirements, and hence rely on tailored material designs specific to the targeted application, limiting their universality. “Reconfigurable memristors” that combine both ionic diffusive and drift mechanisms could address these limitations, but they remain elusive. Here we present a reconfigurable halide perovskite nanocrystal memristor that achieves on-demand switching between diffusive/volatile and drift/non-volatile modes by controllable electrochemical reactions. Judicious selection of the perovskite nanocrystals and organic capping ligands enable state-of-the-art endurance performances in both modes – volatile (2 × 106 cycles) and non-volatile (5.6 × 103 cycles). We demonstrate the relevance of such proof-of-concept perovskite devices on a benchmark reservoir network with volatile recurrent and non-volatile readout layers based on 19,900 measurements across 25 dynamically-configured devices.

The second main advantage of OGB molecule is that the longer carbon chain improves colloidal stability of the NC solution.
It is known that commonly used ligands in the synthesis of lead halide perovskite NCs (oleic acid and oleylamine) loosely bind to the NC surface. This leads to the loss of colloidal stability and structural integrity of perovskite NCs. To overcome instability issues, a new generation of organic ligands with various head groups should be developed and used. CsPbBr3 NCs with quaternary ammonium and novel guanidinium-based long-chain ligands on the surface seem promising in this regard.
The attachment of ligands to the cesium halide terminated surface of perovskite NCs can be imagined as a replacement of some cesium ions by the head groups of the cationic ligands such as DDAB or OGB. It has already been shown that DDAB ligands improve the chemical stability of perovskite NCs and allow their purification preserving high photoluminescence quantum yield 1 . At the same time, we developed a synthesis of OGB-capped CsPbBr3 NCs because we assume that the guanidinium binding group in OGB molecule has a higher ability to create multiple hydrogen bonds with halides on the NC surface in comparison to DDAB molecule that can result in stronger binding to the CsPbBr3 NCs.
Another difference between DDAB and OGB ligands is in their packing density on the surface of perovskite NCs. Carbon chains of DDAB are bulky and stand out which makes their fitting on the NC surface complicated. On the contrary, OGB molecule has only one oleyl chain with a double bond that, most probably, will allow packing of more ligands on the surface of perovskite NCs. DDAB ligand has shorter carbon chains what makes them more useful in applications where charge transfer is important, such as LED devices. However, OGB has a longer carbon chain creating a bigger distance between NCs and making charge transfer harder. All these features described above can strongly impact the performance of devices based on DDAB or OGB-capped CsPbBr3 NCs. within the perovskite matrix also occur at similar energies ~ 0.10 -0.25 eV 3-5 , making it difficult to pinpoint a singular operation mechanism. We hypothesize that during the SET process, Ag atoms are ionized to Ag + and forms a percolation path through the device structure. Electrons from the grounded electrode oxidize

Voltage (V)
Drift behaviour: Upon increasing the Icc to 1mA (3 orders of magnitude higher than that used for volatile threshold switching), permanent and thicker conductive filamentary pathways are possibly formed within the device as illustrated in Supplementary Fig. 6. This increases the device conductance from a high resistance state (HRS) to a permanent and much lower low resistance state (LRS). Electrochemical reactions are triggered to a higher extent and hence, the switching dynamics is now dominated by the drift kinetics of the mobile ion species Ag + and Br − , rather than diffusion. Hence upon removing the electric field, the conductive filaments remain largely unaffected, and the devices retain their LRS and portray longterm plasticity. Application of voltage sweeps, or pulses of opposite polarity causes rupture of these filaments, and the devices are reset to their HRS. For DDAB-capped CsPbBr3 NCs, the devices transition to a non-erasable non-volatile state within ~ 50 cycles, indicating formation of very thick filaments ( Supplementary Fig. 7). On the other hand, the OGB-capped CsPbBr3 NCs display a record-high nonvolatile endurance of 5655 cycles (Fig. 3b) and retention of 10 5 seconds ( Supplementary Fig. 8), pointing to better regulation of the filament formation and rupture kinetics. Supplementary Fig. 6. Proposed non-volatile drift switching mechanism. a-e illustrate the various stages of filament formation and rupture. i-iv indicate the possible reactions happening within the device.

V + I i ii
Thick filament Drift dynamics It is important to note that the abrupt jump in conductance seen in the IV characteristics (    Step Height = 69.4nm

Voltage (V)
This experiment indicates that the resistive switching behaviour in our devices is solely due to Ag filaments.
However, it has to be noted that the nanoscopic electrochemical environment induced by the halide migration during the voltage application could create more favourable pathways for the drift and diffusion of Ag + ions. This has been observed previously in solar cell configurations even across organic transport layers [9][10][11] . Hence, it becomes challenging to completely rule out the effect Brmigration as a possible contributing factor to the resistive switching.

Role of PEDOT:PSS and pTPD interlayers
The PEDOT:PSS + pTPD interlayers allow better perovskite NC thin film formation. The coating of When isolating the effect of the intermixing by normalizing the 107 Ag with respect to a referenced OFF device, the 107 Ag profile of the ON device displays a considerable difference ( Supplementary Fig. 17). The Ag pads deposited on top of the surface seem to be Ag depleted (i.e., a lower amount of Ag), which is confirmed by microscope observation. It must be noted that the first few nanometers of the depth profiling are affected by surface contaminations and can be considered as a transient region. The interesting section is located at the interface between the halide perovskite and the organic layers (near the top red line shown in the plots, also denoted by the circle), where an increase of the 107 Ag count is observed.
Although, the intermixing mentioned above still affects the 107 Ag profile, it is clearly seen by the depth profiles, that there is 107 Ag present in the structure due to the operation of the device. Although the 107 Ag cannot be presently localized precisely in the structure, a qualitative comparison of the Ag profiles between the ON and OFF states proves formation of Ag conductive filaments upon biasing, validating our proposed memristive switching mechanism.  HRS. c shows the calculation of the activation energy by studying the variation in the HRS as a function of temperature.
Dependence on NC layer thickness, size and dispersity Supplementary Fig. 19 shows the device characteristics as a function of the nanocrystal layer thickness. All devices exhibit very similar characteristics with an on-off ratio > 10 3 , similar set and reset voltages.
Statistical analysis of the on-off ratios also reveal independence from the nanocrystal layer thickness.
Supplementary Fig. 19. Effect of NC layer thickness. IV characteristics (top) and statistical analysis of the distribution of on-off ratios during endurance testing (bottom) of 20nm and 90nm OGB-capped CsPbBr3 NC memristors.
To study the effects of NC size and dispersity, 3 solutions of OGB-capped NCs with an average size of 13, 9 and 7nm were prepared. Since NCs have a slightly elongated shape in one direction, it is more convenient to refer to the aspect ratio of the NCs. The corresponding aspect ratios were: 1.47 ± 0.23, 1.26 ± 0.36 and 1.57 ± 0.25 respectively. The DDAB-capped NCs had an average size of 9 nm with an aspect ratio of 1.09 ± 0.1. The size, aspect ratio and monodispersity was confirmed via TEM (not shown here) and photoluminescence measurements (Supplementary Fig. 20). For all colloids, we observed a narrow FWHM of PL spectra (~18 nm for all OGB samples and ~20 nm for DDAB samples) that confirms their high monodispersity.
Supplementary Fig. 20. Photoluminescence spectroscopy of a DDAB and b-d OGB-capped CsPbBr3 NCs of various sizes and aspect ratios in solution.
Devices fabricated with each of the solutions for comparison exhibit very similar characteristics with an on-off ratio > 10 3 , similar set and reset voltages. Statistics derived from extensive measurements also do not show any trend with the nanocrystal size ( Supplementary Fig. 21). While the exact mechanism is still a matter of ongoing research, the best results in terms of yield of devices with an on-off ratio > 10 3 and endurance of > 4000 cycles in the non-volatile mode is obtained with NCs of size 9 nm with an aspect ratio of 1.26 ± 0.36.  solves the classification task with high accuracy without overfitting. b Confidence matrix calculated at the end of training. The correct response probability is shown in the right color scale. It is evident that network performs slightly worse in discriminating irregular patterns. Before the training, we assumed * and " memristors of the readout layer are initialized with RESET, followed by an iterative SET operation, resulting in random conductances in LRS with (0.5, 0.1) mS.

Supplementary
During the inference procedure, reservoir output vector of length 30 is fed into the readout layer.
Memristors in the readout layer are placed in the differential architecture, in which the difference of conductance values of two differential memristors ( * and " ) determines the effective synaptic strength.  For the training procedure, the network loss is calculated as the difference between output layer prediction and the one-hot encoded target vector indicating one of the four firing patterns. At this point, one can calculate targeted weights using the backpropagation algorithm. However, to support fully online-learning, we tested I // controlled weight update scheme where following stages in the pipeline can be easily implemented with the mixed-signal circuits in an event-driven manner. The I // controlled weight update is implemented as follows. First, the required weight change is calculated with 1'2341 = -5 , where is suitably low learning rate, -is the reservoir layer output and 5 is the calculated error. In order to calculate the target conductance values for both positive and negative memristors, we first linearly scale the weight change to conductance change (by multiplying with 1/ ). Secondly, we read both the positive and negative conductance values. By using a push-pull mechanism, we calculate the target conductance values. The push-pull mechanism ensures a higher dynamic range in the differential configuration. Third, the target Thus, it is clear that a memristor that can be reconfigured to achieve on-demand switching between diffusive/volatile and drift/non-volatile modes is a research gap, that we fill with this work.
To demonstrate "reconfigurability" on-the-fly, our devices are switched between volatile and non-volatile modes with precise compliance current (Icc) control and selection of activation voltages. Supplementary   Fig. 29. shows that our devices can act as a volatile memory even after setting it to multiple non-volatile states. This proves true "reconfigurability" of our devices, hitherto undemonstrated. Such behaviour is an example of the neuromorphic implementation of synapses in Spiking Neural Networks (SNNs) that demand both volatile and non-volatile switching properties, simultaneously (see Fig. 1a).
Supplementary Fig. 29. "Reconfigurability" on-the-fly of OGB-capped CsPbBr3 NC memristors. The device is switched between its non-volatile and volatile mode on demand.

Supplementary Note 6: Thermal Camera Imaging
The operation of our perovskite devices depend on the migration of electrochemically active Ag + species through the perovskite matrix as explained in the manuscript. Here, the halide perovskite acts like a scaffold to enable this process. The formation and rupture of such conductive filaments (CFs) depend on a lot of factors. Among these factors, the applied electric field and Joule heating play the most important roles.
Both these factors can play complimentary/competing roles and in determine the nature of the set/reset processes 32,33 .
By setting compliance currents (Icc), we essentially set a limit to the extent of electrochemical reactions that occur during the migration of Ag + species through the perovskite matrix, in turn controlling the thickness of the CF or number of CFs to some extent. When the device is operated under a low Icc of 1µA, the filaments formed are thin and unstable and can dissolve spontaneously, resulting in a volatile memory. And since these processes have a low electrochemical and thermal budget, it becomes feasible to repeat the processes many times, resulting in high volatile endurance 34 . When the Icc is raised to 1mA, the filaments formed are relatively thicker or more number of filaments can be formed, making it difficult to initiate the dissolution process. Hence, the devices preserve the CFs even when powered off, i.e. they are non-volatile.
A large negative voltage is required to reset our bipolar devices and the Joule heating generated during this process ruptures the CFs. In the non-volatile mode, since these processes have a high electrochemical and thermal budget, it becomes difficult to repeat the processes many times, resulting in low non-volatile endurance 35 . This is a common observation in all memristive devices, irrespective of the active switching material as can be seen from Supplementary Note 5 Supplementary Table 2.
To investigate this further, we monitor the working of OGB memristors in their volatile and non-volatile modes in operando using a thermal camera. We observe that maximum heat is generated in our memristors during the reset process when trying to break the conductive filament(s). Application of reverse bias causes Joule heating which ruptures the conductive filament(s). Thick filaments are difficult to break and result in large rise in temperature in the surrounding areas which we pick up as an infrared image [36][37][38] .
Supplementary Fig. 30 shows the thermograms recorded at the endurance limit of the respective modes -2 million cycles for the volatile and 5000 cycles for the non-volatile mode. When the OGB device reaches its maximum non-volatile endurance/near failure (5000 cycles), an 84 o C rise in surface temperature is noted. These findings indicate that an applied field alone cannot rupture the conductive filaments and that the reset process is likely to be a combined effect of electric field and Joule heating. However, no rise in surface temperature is observable after 2 million cycles of volatile endurance, supporting our hypothesis of better management of the electrochemical reactions with lower Icc. As a result, the volatile endurance is much larger when compared to the non-volatile endurance.
Supplementary Fig. 30. In operando monitoring of OGB memristors in their volatile and non-volatile modes using an infrared camera.
The dynamic ligand binding in halide perovskite (HP) nanocrystals (NCs) have been widely observed to create surface traps 39 . Therefore, handling the surface ligands without sacrificing the optoelectronic properties or affecting the structural integrity poses a challenge to developing semiconductive HP NC films.
The large PL quenching observed with DDAB ligands (solution vs thin film comparison shown in Supplementary Fig. 12) point to creation of large number of surface traps. In perovskite films and NCs, it has been firmly established that trap states at the grain boundaries and on the surface can capture photoexcited charge carriers to create a local electric field capable of promoting ion migration [40][41][42][43][44][45] . In our case, this could lead to enhanced Ag + and Brmigration and eventually lead to thicker Ag filament formation.
From our IV endurance measurements, the DDAB devices are observed to (i) quickly transit from a volatile to a non-volatile state, even at a low compliance current (Icc) of 1 µA resulting in an inferior volatile endurance of ~ 10 cycles, and (ii) quickly transit to a non-erasable non-volatile state at high Icc of 1 mA, resulting in an inferior nonvolatile endurance of ~ 50 cycles.
Both these results support the hypothesis of enhanced electrochemical reactions in the DDAB system due to their short chains.
To investigate this further, we monitor the working of DDAB and OGB memristors in their non-volatile modes in operando using a thermal camera. Supplementary Fig. 31 shows the thermograms recorded at the time of failure of the memristors in their non-volatile mode-50 cycles for DDAB and 5655 cycles for OGB.
For the OGB device, a thermogram is also recorded at the 50 th endurance cycle for direct comparison to the DDAB device. At the time of failure of the DDAB device, i.e. 50 th endurance cycle, a 95 o C rise in surface temperature is seen. In comparison, only a 13 o C rise is observed for OGB devices at the 50 th cycle. When the OGB device reaches its maximum endurance/near failure (5655 cycles), a 86 o C rise in surface temperature is noted similar to the DDAB device. These findings indicate that the reset process is likely to be a combined effect of electric field and Joule heating. Hence it is critical to engineer materials to regulate the underlying electrochemical reactions. This experiment further supports our hypothesis of better management of the electrochemical reactions with the OGB ligands when compared to DDAB, and points to the importance of investigating nanocrystal-ligand chemistry for the development of high-performance robust memristors.

Supplementary Note 7: Thin film vs NC film
To compare thin film and NC film of CsPbBr3, the thickness of the NC layer is increased to ~100nm by 5- Most importantly, we observe that CsPbBr3 thin film memristors fail to be reconfigured back to their volatile mode once their non-volatile mode is activated, similar to other dual functional memristors reported in literature. On the other hand, the NC-based memristor allow facile reconfiguration between the volatile and non-volatile modes ( Supplementary Fig. 33).