betaFeedback

WorksPhysics and Astronomy

Tunable intervalence charge transfer in ruthenium Prussian blue analogue enables stable and efficient biocompatible artificial synapses

Abstract

Emerging concepts for neuromorphic computing, bioelectronics, and brain-computer interfacing inspire new research avenues aimed at understanding the relationship between oxidation state and conductivity in unexplored materials. Here, we present ruthenium Prussian blue analogue (RuPBA), a mixed valence coordination compound with an open framework structure and ability to conduct both ionic and electronic charge, for flexible artificial synapses that reversibly switch conductance by more than four orders of magnitude based on electrochemically tunable oxidation state. Retention of programmed states is improved by nearly two orders of magnitude compared to the extensively studied organic polymers, thus reducing the frequency, complexity and energy costs associated with error correction schemes. We demonstrate dopamine detection using RuPBA synapses and biocompatibility with neuronal cells, evoking prospective application for brain-computer interfacing. By application of electron transfer theory to in-situ spectroscopic probing of intervalence charge transfer, we elucidate a switching mechanism whereby the degree of mixed valency between N-coordinated Ru sites controls the carrier concentration and mobility, as supported by DFT.

Mixed valence compounds (MVCs) that accommodate facile ion insertion and extraction have recently garnered intense interest for applications beyond energy storage such as neuromorphic computing, thermal and magnetic switching, and electrocatalysis.[1] This interest is motivated by the possibility to tune the electronic structure of the host via the coupled ion/electron insertion/extraction reactions characteristic of electrochemical charge transfer (CT). Among the various classes of MVCs, the continually growing family of Prussian blue analog (PBA) coordination frameworks[2, 3, 4] stand out due to their high (electro)chemical stability,[5, 6, 7] fast ionic transport,[8, 9] biocompatibility and applications for biosensing,[10, 11] process compatibility with electron and photolithography,[12, 13] high electrocatalytic activity,[5, 6] and synthetic tunability.[3, 4] However, fundamental insight into the electronic conduction mechanism remains limited for most PBAs,[14, 15] including the archetypal PB,[16, 17, 18, 19] thus hindering their utilization for electronics applications. This is surprising, given the success of electron transfer theory in explaining the electronic properties of many mixed valence coordination compounds,[20, 21, 22] including cyanide-bridged complexes that mimic the molecular building blocks of PBAs.[23, 24, 25]

In this report, we describe a Ru-based Prussian blue analogue[26] (RuPBA) as both a model system for understanding redox-tunable electronic delocalization in extended mixed valence coordination compounds[27, 28] and as an efficient active material for flexible electronic devices that rival organic electrochemical transistors (OECTs)[29, 30] for emerging analog memory[31, 32, 33] and biointerfacing[34] applications. We elucidate the interrelationship between the electronic conductivity and the electrochemically tunable rate of intervalence CT using Marcus-Hush theory (MHT),[35, 36] Boltzmann transport theory, and operando absorption spectroscopy, indicating that hole mobility >1 cm2/(V⋅s)>1\ \mathrm{cm}^{2}/(\mathrm{V}\cdot\mathrm{s}) is achievable in this material. Application of MHT to the absorption spectra at different oxidation states reveals that electronic conduction predominantly occurs through a cross-pore (CP) charge transfer pathway between N-coordinated Ru sites, as depicted in Figure 1a, consistent with DFT predictions. To show the potential of RuPBA in neuromorphic computing, we demonstrate electrochemical random-access memory (ECRAM) artificial synapses[31, 37, 38] that switch conductance over a range of >4>4 orders of magnitude with a large linear regime, low read/write currents, strong endurance, and excellent long-term state retention. The RuPBA ECRAM is inkjet printed on flexible substrates, biocompatible, and sensitive to dopamine. To our knowledge, the RuPBA represents the first practical alternative material to polymer blend based OECTs, offering promising avenues for innovation via its distinct ion/electron conduction mechanism and large synthetic flexibility of the PBA family of compounds.

Figure 1. Electrochemical doping of RuPBA. a, Idealized vacancy-free structure of partially lithiated RuPBA, Li3Ru2[Ru(CN)6]2\mathrm{Li}_{3}\mathrm{Ru}_{2}[\mathrm{Ru}(\mathrm{CN})_{6}]_{2}, showing proposed cross-pore (CPCP) and nearest neighbor (NNNN) charge transfer paths. b, Schematic representing reversible Li+\mathrm{Li}^{+} ion insertion into the RuPBA film at the gate and channel separated by a Li+\mathrm{Li}^{+}-containing ionogel electrolyte (See Supplementary Information, section S2, for a detailed description of the ECRAM device architecture). c, Transfer curve for a RuPBA/Li-ionogel device showing the resulting source-drain current of the channel (isdi_{\mathrm{sd}}) on a linear scale (left y-axis, blue) and channel conductance on a logarithmic scale (GG, right y-axis, red) when scanning the gate-drain voltage (VgdV_{\mathrm{gd}}) at 30 μV/s30\ \mu\mathrm{V}/\mathrm{s} and applying a source-drain bias of 0.1 V0.1\ \mathrm{V}. The dashed box designates the region where isdi_{\mathrm{sd}} increases linearly as a function of electrochemical charge. d, Scheme conveying the change in Ru valency and degree of lithiation as the channel undergoes oxidation (Ox) and reduction (Re). Numeric labels on Ru atoms represent formal oxidation states and valence mixing. The region shown in pink highlights the proposed cross-pore mixed valence conduction pathway, where electron delocalization occurs between neighboring N-coordinated Ru centers while C-coordinated Ru centers maintain localized 2+2+ formal oxidation states.

Figure 1a shows the idealized structure of RuPBA, depicted without [Ru(CN)6]4−[\mathrm{Ru}(\mathrm{CN})_6]^{4-} vacancies for simplicity, containing lithium ions intercalated within the square faces[39] of the cubic CN−\mathrm{CN}^{-}-bridged cages. The as-synthesized RuPBA contains 3 at.% of K+\mathrm{K}^{+} ions (Ru:K = 2.5:1), indicating that most of the Ru centers initially have a 2+2+ formal oxidation state. We expect the K+\mathrm{K}^{+} ions exchange with the Li+\mathrm{Li}^{+} ions after the silica-based Li+\mathrm{Li}^{+}-containing ionogel electrolyte[40] layer is printed over the RuPBA layer. The transfer curve in Figure 1c shows electrochemical tuning of the channel conductance from a lower limit of ∼9×10−9\sim9 \times10^{-9} S to an upper limit of ∼2×10−4\sim2 \times10^{-4} S. Based on the RuPBA film conductivity in the as deposited state (4×10−44 \times10^{-4} S/cm, Extended Data Figure 1) and the measured conductance from Figure 1c (2×10−52 \times10^{-5} S at Vgd=0V_{\mathrm{gd}} = 0 V), the electrochemically tunable conductivity range extends from ca. 10−610^{-6} to 10−210^{-2} S/cm. Interestingly, the lower conductivity at the fully reduced RuPBA (fr-RuPBA) state (∼10−6\sim10^{-6} S/cm) is close to that obtained by Xidis and Neff for fr-PB (5×10−75 \times10^{-7} S/cm).[18] We attribute the lower conductance limit at Vgd∼0.6V_{\mathrm{gd}} \sim0.6 V to that of fr-RuPBA, where all Ru centers have a 2+2+ formal oxidation state and the structure contains the maximum amount of lithium ions. Partial oxidation of fr-RuPBA causes RuII/III\mathrm{Ru}^{\mathrm{II/III}} mixed valency and conductivity to increase.

Extended Data Figure 1. DC conductivity-temperature relationship for RuPBA at the as-synthesized valence state. a, Log(conductivity, σ\sigma) vs reciprocal temperature with corresponding activation energy (EaE_{\mathrm{a}}) labelled on plot. Inset image shows the inkjet-printed 2×22 \times2 mm RuPBA film on a Au/PET van der Pauw square design, film thickness of 0.6±0.10.6 \pm0.1 μ\mum. b, Current-voltage recordings between two adjacent terminals of the van der Pauw sample at varied temperature.

Most of the accumulated electrochemical gating charge should be proportional to the amount of RuII\mathrm{Ru}^{\mathrm{II}} species oxidized according to Faraday’s law, and thus should also scale linearly with the charge carrier concentration. Analysis of the transfer data shows a voltage range where the conductance changes linearly with accumulated gate charge (see also Supplementary Information, Section S1). The trend suggests that only carrier concentration changes in the linear region, while mobility remains constant. Outside of this region, oxidation of the RuPBA channel either causes the mobility to increase (0.60.6 V ≲Vgd≲−0.1\lesssim V_{\mathrm{gd}} \lesssim-0.1 V) or decrease (−0.3-0.3 V ≲Vgd≲−0.6\lesssim V_{\mathrm{gd}} \lesssim-0.6 V). Upon oxidizing the channel further, the conductance decreases, indicating that a mixed valence form is the most conductive (Extended Data Figure 2).

Extended Data Figure 2. Transfer curve for a RuPBA/Li-ionogel device when the gate-drain voltage is scanned to -1 V at 30 μV/s, resulting in more partial oxidation of RuPBA in the channel in comparison to Figure 1c. Source-drain bias of 0.1 V.

Evaluation of ECRAM device performance

Consistent with the transfer characteristics (Figure 1c), the voltage-controlled pulsed programming of an ECRAM device with a 0.75 ms0.75\ \mathrm{ms} ‘write’ pulse duration shows that the potentiation (G↑G\uparrow) and depression (G↓G\downarrow) ramps (Figure 2a) are highly linear, which is desirable for high accuracy during training.[32, 33] Devices fabricated with a proton-conducting polymer-based ionogel electrolyte show better switching behavior in terms of linearity and symmetry, but switch more slowly, requiring more than 10-fold longer write pulse durations and twice as many pulses to cycle over a range of Gmax⁡=2Gmin⁡G_{\max} = 2G_{\min} (Extended Data Figure 3). The conductance range of 80–160 nS in Figure 2a corresponds to a read current range of 8–16 nA at the applied read voltage (Vr=0.1V_{\mathrm{r}} = 0.1 V). The combination of low current and near-linear conductance tuning is ideal for energy-efficient programming.[32]

Li-RuPBA ECRAM performance

Figure 2. Li-RuPBA ECRAM performance. a, Channel conductance during pulsed programming at Vgd=+1V_{\mathrm{gd}} = +1 V/-1 V for depression/potentiation at 0.75 ms gate pulse duration. b, Long term retention of 5 programmed conductance states. c, Simulated performance for neuromorphic tasks. Error bands represent standard deviation. d, Pulsed conductance programming with different channel dimensions, as labelled. ‘Write Time’ represents the total time that voltage is applied to the gate. Log-log plots showing relationship of channel area, AA, to e, switching times, Δtmin⁡\Delta t_{\min}, and f, switching energies, ΔEmin⁡\Delta E_{\min}, required to change the channel conductance by an average value of ±1.0\pm1.0 nS. Inset tables

Extended Data Figure 3. Channel conductance during pulsed programming at Vgd=+1V_{\mathrm{gd}} = +1 V/−1-1 V for depression/potentiation at 10 ms gate pulse durations for an RuPBA device employing a proton-conducting ionogel electrolyte.

State retention is essential for nonvolatile synaptic memory. As shown in Figure 2b, the programmed conductance over the entire range, 0.65–2.0 μ\muS (Gmax⁡/Gmin⁡∼3G_{\max}/G_{\min} \sim3), has a retention time of ∼\sim600 s (defined as dG<1%dG < 1\%). The loss in retention at G>G0G > G_{0} in Figure 2b indicates that the more oxidized RuPBA states are less stable than the reduced states. However, the conductance remains stable within ±1%\pm1\% over a duration of ∼104\sim10^{4} s for the range below the initial state, 0.65–1.5 μ\muS (Gmax⁡/Gmin⁡∼2G_{\max}/G_{\min} \sim2), an improvement in comparison to the ∼\sim250 s retention time reported for PEDOT:PSS synapses programmed within a similar Gmax⁡/Gmin⁡G_{\max}/G_{\min} range under inert environment.[41] PEDOT:PSS channels contain polyethyleneimine (PEI), which acts as an electron donor to annihilate electrochemically generated holes until the conductivity decays to its initial state. When charged to a lower conductivity, the PEI additive diffuses out of the channel and into the electrolyte, causing conductivity to increase over time towards the initial state.[41, 42] Given standard neural network training, the negative impacts of such volatility on network accuracy and energy efficiency might be considerable, e.g., a deep neural network based on filamentary ReRAM where a ∼\sim5% conductance change over a period of 10410^{4} s at elevated temperature led to ∼\sim16% loss in network accuracy, and could only be mitigated through a correction scheme applied every ∼103\sim10^{3} s.[43, 44] Yet, the conductance of RuPBA, when programmed at the lower range (G<G0G < G_{0}), does not trend back towards G0G_{0} because RuPBA does not rely on a redox buffer such as PEI. This implies that, with only mild loss of analog range, non-volatile programming in the range of 0.65–1.5 μ\muS can still enable efficient in-situ learning and retention for embedded biocompatible applications, with only minor alterations to normal training needed (e.g., an array calibration step after each epoch to detect outlier weights exceeding some threshold above G0G_{0}, where they can be clipped/RESET). As most neural network weights are low significance (low-conductance), the hardware mapping for this scheme is natural and could lead to low overall error profiles at inference stage.[45] Meanwhile, the range clipping would not be a major detriment to learning performance, since a compressed range of 6-8 bits writable space is more than sufficient for most online learning applications using emerging non-volatile memory devices[46, 47] even when considering write noise in the loop.[48]

show values obtained by extrapolation of the linear fits (red lines). The durations of the applied write pulses were 0.75 ms, 6 ms, and 10 ms in order of increasing area; this was done to achieve roughly equal numbers of conductance states per ramp, ~70 states for each of the three devices. The channel length:width (L:W) aspect ratio of devices for parts d-f is 1:7.5, keeping the RuPBA film thickness constant at ~0.6 μm. Error bars represent the standard deviations for parts e-f. See Table S3 of the Supplementary Information for number of switching pulses analyzed per device. All measurements were performed under N2_2 atmosphere at T = 300 K.

LL / μm\mu\mathrm{m}Δtw\Delta t_{\mathrm{w}} / msTotal pulsesStates per rampΔGavg\Delta G_{\mathrm{avg}} / nSσΔG\sigma_{\Delta G} / nSΔGmin⁡\Delta G_{\min} / nSΔtmin⁡\Delta t_{\min} / msΔQavg\Delta Q_{\mathrm{avg}} / nCΔQmin⁡\Delta Q_{\min} / nCΔEmin⁡\Delta E_{\min} / nJ
1000.7523569831.00.510.7 ±\pm0.41.51 ±\pm0.020.8 ±\pm0.40.8 ±\pm0.4
3006.009528641.30.415 ±\pm116 ±\pm16 ±\pm26 ±\pm2
40010.008265731.10.619 ±\pm528 ±\pm113 ±\pm613 ±\pm6
Mean731.10.5
Std. dev.90.10.1

Table S3. Measured Parameters Used to Determine Minimum Switching Time and Energy

The channel conductance can be repeatedly cycled using 10610^{6} switching pulses without altering switching characteristics (Extended Data Figure S4). From this data, we evaluate online learning by simulating a hardware multi-layer perceptron populated with RuBPA synapses. Due to the existence of a smooth, linear workable region optimal for on-chip backpropagation,[49] the RuBPA synapse generally performs excellently in the context of online learning. As shown in Figure 2c, on smaller tasks, such as the Iris flower classification problem[50] and the optical character recognition (OCR) task,[51, 52] RuPBA synapses achieve close agreement with ideal synapses (numerical). On larger neuromorphic tasks such as the fashion-MNIST[53] and MNIST task,[54] the gap between ideal and RuPBA devices diverges slightly more. Overall, our results are comparable to those of other leading linear and symmetric nonvolatile memory candidates (Supplementary Information, Section S3).[31, 33, 37, 55, 56, 57]

Figure S4. RuPBA ECRAM dopamine sensor operation. a, Recordings of channel conductance (GG), gate voltage (VgV_{\mathrm{g}}), and gate current (igi_{\mathrm{g}}) during incremental additions of DA. Black sections of the VgV_{\mathrm{g}} trace show when the voltage was measured at open-circuit and red points mark when short voltage pulses were applied to the gate (p1p_1: 0.0 V, 100 ms, p2p_2: -0.1 V, 435 ms). b, Schematic depicting how the electrochemical potential and RuII\mathrm{Ru}^{\mathrm{II}} concentration change at the gate and channel at different stages of operation. The steps in part b are labeled in the zoomed-in section of the VgV_{\mathrm{g}} trace of part a for the first addition of DA solution. VsdV_{\mathrm{sd}} = 0.1 V.

The programming speed and energy efficiency of an ECRAM device depends on the electrochemical current of the ‘write’ step when a gate voltage is applied to change the redox state of the channel. Write currents should be minimized to reduce the energy cost.[31] However, higher faradaic write current densities should increase the ECRAM switching rate (faster charging). Figure 2d shows conductance vs. write time traces for three devices with different channel areas, AA, but a similar length:width aspect ratio (L:W=1:7.5L:W = 1:7.5). The smallest device clearly operates more quickly, completing 12 potentiation/depression cycles before the larger device can finish the first cycle. The average times and energies required to switch the conductance by a magnitude of 1 nS are plotted versus channel area in Figure 5e-Figure 5f (See Supplementary Information, section S4). The smallest device achieves a switching time of 0.75 ms, comparable to the duration of action potentials in biological neurons.[58] The relationships are highly linear with log-log slopes of 1. From extrapolation, we predict that RuPBA ECRAM devices of similar architecture with 100 nm channel length should switch on the timescale of ~1 ns, much faster than biosynapses, with ~1 fJ write energy, a lower energy cost compared to ~10 fJ per synaptic event in biological synapses.[58] We anticipate that RuPBA devices fabricated with thinly layered channel/electrolyte/gate vertical architecture[37] would significantly lower the electrolytic gating resistance and increase the overall switching efficiency.[32]

Figure 5. Interpretation of optical charge transfer bands from UV/Vis/NIR absorption spectrum of RuPBA. a, Reduced absorption spectrum and fitted asymmetric gaussian peaks, highlighting two different IVCT bands for transfer between nearest neighbor cyanide bridged Ru centers (IVCTNN_{\mathrm{NN}}) and cross-pore between N-coordinated Ru centers (IVCTCP_{\mathrm{CP}}), as depicted in the inset scheme. The absorbance is divided by ν\nu to accurately determine the reorganization energy and coupling in accord with theory.[22] b, Schematic representation for potential energy surface diagram of IVCT based on a three-state model. Red dashed lines labeled A, B, and C convey three parabolic overlapping diabatic energy surfaces, where the energy minima for A,B are equivalent and occur at X=0,1X = 0,1. The higher energy minimum for C occurs at X=0.5X = 0.5. Blue lines labeled 1, 2, and 3 represent the adiabatic states resulting from electronic coupling between N-coordinated donors, acceptors, and hexacyanoruthenate bridge states. The

Potential for biosensing and brain-computer interfacing

RuPBA undergoes reduction upon reaction with dopamine (DA), an important neurotransmitter. Figure 3a shows the design for a modified ECRAM device that allows DA to react with the RuPBA gate while preventing direct reaction with the channel. Figure 3d shows that the device conductance decreases by ΔG\Delta G when incrementally increasing the DA concentration, leading to the dose-response calibration curve. (see Supplementary Information, § 14, for a more detailed description of sensor operation). The RuPBA synapse achieves a 4 μM4\,\mu\mathrm{M} limit of detection, comparable to PEDOT:PSS-based OECTs.[34, 59]

Figure 3. Dopamine (DA) sensing with modified ECRAM cell. a, Sketch of modified device for dopamine sensing. b, Illustration of dopamine release from vesicles of a presynaptic neuron (top) to receptors of a postsynaptic neuron (bottom). c, Schematic of dopamine reaction with RuPBA in the portion of the device gate enclosed by the well. d, Conductance trace during incremental additions of dopamine and inset showing corresponding dose-response plot.

In the inset, ∣ΔG∣=∣G−G0∣|\Delta G| = |G - G_{0}|, where G0G_{0} (= 4.85 μ\muS) is indicated by the topmost dashed horizontal red line on the conductance trace.

The sensitivity of RuPBA to dopamine establishes a basis to develop hybrid bio/artificial synapses.[34] For successful bio-interfacing, neuronal cells should be able to grow and thrive on RuPBA films. We find that cells seeded onto RuPBA films divide at a rate comparable to control cells, and that their viability is as high as control cells (Extended Data Figure 5). Optical and scanning electron micrographs of the cells show interaction with and adherence to the RuPBA substrate (Extended Data Figure 5-Extended Data Figure 5). We conclude that RuPBA is nontoxic and nonrefractory to cell growth. The compatibility of RuPBA with live neuronal cells signifies a promising first step towards utilizing electrochemical transistors to integrate biological and artificial analog neural networks.

Extended Data Figure 5. Compatibility of neuronal cells grown on RuPBA films. a, Growth and viability assays for Neuro2A cells grown on well plates modified with or without RuPBA films (blue or red columns, respectively). Left panel shows cell count after 3 days of incubation, each well initially seeded with 10,000, 25,000 or 50,000 cells, and right panel shows corresponding percentage of viable cells (percent live) based on Trypan Blue staining. Error bars represent the standard error of the mean based on 4 replicates. b-d, Optical micrographs and e, SEM image of Neuro2A cells grown on RuPBA films.

Spectroscopic probing of RuPBA electronic structure during switching

In-situ UV/Vis/NIR spectroscopy provides a means to probe the electronic structure of RuPBA during electrochemical cycling. The visible color change of an electrochromic RuPBA film on FTO/glass is shown in the Supplementary Information, Video S1. The film changes color from beige at the more reduced state to blue grey at the more oxidized state. Figure 4a shows the evolution of absorption spectra as a RuPBA/ITO electrode is scanned from a fully reduced state at -0.2 V to a partially oxidized state at 1.0 V (versus Ag/AgCl wire). Following Behera et al., we assign the bands at ca. 30,000 cm−1^{-1} and 6000 cm−1^{-1} to the metal-to-ligand charge transfer (MLCT) and intervalence charge transfer (IVCT) reactions, respectively.[26] As the film is oxidized, the MLCT intensity decreases while the IVCT intensity increases. Like the conductance vs gate voltage trend of Figure 1c, the relationship of IVCT absorbance to electrode potential in Figure 4b is sigmoidal and reversible. The absorbance vs potential response is consistent over multiple cycles in this 1.2 V potential window. Figure 4c shows the interrelationship of applied gate voltage, IVCT absorbance, and conductance when switching conductance across four orders of magnitude in a RuPBA ECRAM device. The correlation of conductance to IVCT absorbance signifies that the electronic conductivity of RuPBA is inextricably linked to the rate of intervalence charge transfer.

In-situ UV-Vis-NIR absorption spectroscopy of RuPBA films during voltammetric cycling

Figure 4. In-situ UV-Vis-NIR absorption spectroscopy of RuPBA films during voltammetric cycling. a-b, Results from cyclic voltammetry of RuPBA film on indium tin oxide (ITO)-coated glass in contact with liquid electrolyte containing 1.0 M LiFSI dissolved in Pyr14_{14}TFSI, scan rate of 50 μV/s. a, Overlaid spectra recorded as the potential, EE, was scanned from -0.2 to 1.0 V (vs Ag/AgCl wire quasireference electrode). b, Absorbance monitored at 6452 cm−1^{-1}, as taken from the 2nd scan cycle, with the initially recorded open circuit potential at equilibrium, ‘initial state’, marked by the gray dashed lines. c, Measurements of the conductance and IVCT absorbance in the channel region of an ECRAM device after potentiostatic charging at varied gate voltages (VgdV_{\mathrm{gd}}). Conductance is shown on a logarithmic scale. Inset image on the conductance plot shows the optical path from the tungsten halogen light source through the device channel.

When extending the applied potential to 1.4 V, we observe an additional trend in the IVCT region of the UV/Vis/NIR spectra (Extended Data Figure 6). The lowest energy maximum at ca. 6450 cm−1^{-1} follows the trend as discussed above from -0.2 to 1.0 V, but then decreases in absorbance from 1.0 to 1.4 V. The absorbance at ca. 13,000 cm−1^{-1}, however, continues to increase as RuPBA undergoes oxidation to 1.4 V. The two observed trends reveal that two distinct IVCT processes occur in RuPBA, and that the absorbance of the lower energy IVCT correlates with conductivity.

Extended Data Figure 6. UV/Vis/NIR spectroscopic monitoring of redox processes for a RuPBA film on ITO-coated glass electrode in a cell containing 1.0 M LiFSI dissolved in Pyr14TFSI\mathrm{Pyr}_{14}\mathrm{TFSI} ionic liquid. a, Overlaid absorption spectra recorded while scanning the potential from 1.0 to 1.4 V. Current-potential response from cyclic voltammetry at 50 μ\muV/s, starting at -0.2 V (second cycle). Vertical dashed line marks the OCV of the initial state (315 mV) prior to performing the voltammetric scans. Overlaid absorption spectra recorded while scanning the potential from b, -0.2 to 1.0 V, and c, 1.0 to 1.4 V. d, Absorbance vs potential as recorded at three different wavelengths during the scan. Potential is reported versus Ag/AgCl wire reference electrode.

Mechanistic interpretation

The strong correlation of IVCT absorbance with conductivity supports a mechanism whereby electronic transport in RuPBA occurs through adiabatic intramolecular CT reactions. As shown in Figure 5a, deconvolution of the UV/Vis/NIR spectrum for a film of the as-synthesized RuPBA (as-RuPBA) suggests at least seven different overlapping optical absorption bands in the wavenumber (ν\nu) range of 35,000 to 4,000 cm−1^{-1}, consistent with the previous report.[26] We assign the two lowest energy bands at ca. 8400 cm−1^{-1} (1.0 eV) and 5100 cm−1^{-1} (0.6 eV) to two different IVCT transitions, as depicted in Figure 5a. The higher energy band (IVCTNN_{\mathrm{NN}}) corresponds to electron transfer from N-coordinated Ru (N-Ru) to its nearest neighbor, C-coordinated Ru (C-Ru), as separated by the cyanide bridge. The lower energy band (IVCTCP_{\mathrm{CP}}) corresponds to ‘cross-pore’ transfer between two N-coordinated RuII^{\mathrm{II}}/RuIII^{\mathrm{III}} redox pairs. Our interpretation of these two separate IVCT bands agrees with that of isolated trinuclear and tetranuclear CN-bridged Ru complexes.[23, 60, 61, 62] Both IVCT bands are best fit using an asymmetric peak shape, which indicates that the as-RuPBA fits in a transition regime of electronic delocalization according to Robin-Day classification,[63] exhibiting properties typical for both class 2 (semiconductor) and class 3 (metallic) systems, or the so-called “class 2B” system.[22] In contrast, PB belongs to the class 2A regime involving more localized charges.[63]

IVCTCP\mathrm{IVCT}_{\mathrm{CP}} product and reactant states are of equivalent ground state energy due to symmetry. The third possible state (state 3), formed from IVCTNN\mathrm{IVCT}_{\mathrm{NN}}, is thermodynamically disfavored. EopNNE_{\mathrm{op}}^{\mathrm{NN}} and EopCPE_{\mathrm{op}}^{\mathrm{CP}} represent hνh\nu for the deconvoluted peak maxima of IVCTNN\mathrm{IVCT}_{\mathrm{NN}} and IVCTCP\mathrm{IVCT}_{\mathrm{CP}}, respectively.

The effective donor-acceptor coupling, HeffH_{\mathrm{eff}}, and activation energy, Ea∗E_{\mathrm{a}}^{*}, were determined based on the IVCTCP\mathrm{IVCT}_{\mathrm{CP}} peak energy and shape following an equivalent two-state model[22] for the three-state system depicted in Figure 5b (see Supplementary Information, § 15). The fit of the IVCTCP\mathrm{IVCT}_{\mathrm{CP}} band for as-RuPBA gives EopCP=λ=0.64E_{\mathrm{op}}^{\mathrm{CP}} = \lambda= 0.64 eV and a large coupling, Heff=0.22H_{\mathrm{eff}} = 0.22 eV; this leads to Ea∗=16E_{\mathrm{a}}^{*} = 16 meV, which is less than kBTk_{\mathrm{B}}T at room temperature. The IVCTCP\mathrm{IVCT}_{\mathrm{CP}} rate constant, kctk_{\mathrm{ct}}, is then

kct=κνnexp⁡(−Ea∗/kBT)(1)k_{\mathrm{ct}} = \kappa\nu_{\mathrm{n}}\exp(-E_{\mathrm{a}}^{*}/k_{\mathrm{B}}T) \tag*{(1)}

where κ\kappa is the probability of electron transfer at the configuration coordinates (XX) when the activation barrier is minimized and νn\nu_{\mathrm{n}} is a nuclear vibration frequency,[21] here assigned as the cyanide bridging ligand vibration (ca. 2130 cm−1^{-1}) shown in Extended Data Figure 7, giving νn=6.4×1013\nu_{\mathrm{n}} = 6.4 \times10^{13} s−1^{-1}. The calculated electronic frequency[20, 64] (νel=1.0×1015\nu_{\mathrm{el}} = 1.0 \times10^{15} s−1^{-1}, see Supplementary Information, § 15) is much higher than νn\nu_{\mathrm{n}}, such that κ\kappa is essentially unity, as expected. This leads to a very fast rate constant of kct≈3×1013k_{\mathrm{ct}} \approx3 \times10^{13} s−1^{-1}, only slightly lower than νn\nu_{\mathrm{n}}, further supporting a Robin-Day class 2B assignment for as-RuPBA, but very close to class 3 delocalization.[22] The kctk_{\mathrm{ct}} value corresponds to an IVCT timescale of 30 fs, which is within the sub-picosecond range measured for CT in PBAs by time-resolved spectroscopy.[65, 66, 67]

Extended Data Figure 7. In-situ Raman spectroscopy of ECRAM device channel. a, Raman spectra of cyanide stretch region recorded in the device channel after charging for 30 minutes at 3 different gate voltages, as labeled on the plot. Vertical lines drawn as a guide for the peak maxima for the most oxidized channel at Vgd=−0.5V_{\mathrm{gd}} = -0.5 V (red dotted line) and the most reduced channel at Vgd=+0.5V_{\mathrm{gd}} = +0.5 V (blue dotted line). b, Conductance traces (log scale) recorded during Raman scan after gate charging.

Rosseinsky and coworkers presented a simple expression for IVCT-based conductivity of PB following the treatments described by Austin-Mott and Robin-Day.[63, 68, 69] We propose a similar relationship below.

σ=pe2d2kct/2kBT=epμh(2)\sigma= pe^{2}d^{2}k_{\mathrm{ct}}/2k_{\mathrm{B}}T = ep\mu_{\mathrm{h}} \tag*{(2)}

Here, the conductivity (σ\sigma) depends on the number density of hole charge carriers (pp) and the distance between reactant centers (dd). For IVCTCP\mathrm{IVCT}_{\mathrm{CP}}, the shortest distance between N–Ru centers is d≈7.4d \approx7.4 Å based on the XRD pattern of RuPBA (Extended Data Figure 8). Eq. (2) also shows how kctk_{\mathrm{ct}} and dd relate to the hole mobility; μh=ed2kct/2kBT\mu_{\mathrm{h}} = ed^{2}k_{\mathrm{ct}}/2k_{\mathrm{B}}T. For as-RuPBA, μh≈4\mu_{\mathrm{h}} \approx4 cm2^{2}/(V⋅s\mathrm{V}\cdot\mathrm{s}), which is comparable to that reported for PEDOT:PSS.[70] Because both IVCTCP\mathrm{IVCT}_{\mathrm{CP}} absorbance and conductivity increase as the RuPBA partially oxidizes, we propose that pp must relate to the concentration of mixed valence N-Ru sites, as depicted in Figure 5. The maximum carrier concentration occurs when the valence mixing is 1:1 II:III between N-Ru centers, corresponding to a net mixing of 3:1 II:III for all Ru centers, assuming an ideal defect-free lattice.

Extended Data Figure 8. Extended Data Fig. 8. Structure and morphology of RuPBA. a, Observed powder x-ray diffraction pattern (blue) and calculated[82] pattern (red), assuming face-centered cubic symmetry and a lattice parameter of 10.4 Å, with Miller indices labeled on the plot for the dominant lattice planes. b, Scanning electron micrograph of RuPBA.

Closer analysis of the spectra in Figure 4a reveals that the IVCT band manifold not only increases in absorbance, but also shifts by more than 0.3 eV to lower energy throughout the course of oxidation from fr-RuPBA to the most conductive mixed valence state at 1.0 V, signifying that λ\lambda and Ea∗E_{\mathrm{a}}^{*} decrease while kctk_{\mathrm{ct}} and μh\mu_{\mathrm{h}} increase (see Supplementary Information, Section S6). Spectroscopically determined parameters at varied EE (Extended Data Table S1) indicate that mobility increases from 1 to 4 cm2^{2}/(V⋅s\mathrm{V}\cdot\mathrm{s}) when oxidizing RuPBA over a small potential range from E=0E = 0 to 0.1 V. However, the mobility does not vary significantly at potentials between 0.1 V and the most conductive state (E=1.0E = 1.0 V), implying that the increasing carrier concentration from oxidation is the sole contributor to conductance switching in the VgdV_{\mathrm{gd}} range roughly between 0 and -0.5 V (Figure 1c). The trend of increasing, constant, then decreasing mobility during partial oxidation of fr-RuPBA agrees well with the interpretation of the transfer curve above (Figure 1c and Supplementary Information, Figure S1).

Relationship between source-drain current and electrochemical charge during RuPBA channel oxidation

Figure S1. Relationship between source-drain current and electrochemical charge (channel oxidation charge) during the oxidation of the RuPBA channel from Vgd=0.6 VV_{\mathrm{gd}} = 0.6\ \mathrm{V} to −0.6 V-0.6\ \mathrm{V} as shown in Figure 1c of the main text. The range of RuPBA oxidation states where isdi_{\mathrm{sd}} increases linearly with accumulated charge is encompassed within the vertical dashed lines and the corresponding VgdV_{\mathrm{gd}} range is highlighted by the horizontal red dashed lines. The purple arrows indicate gate voltage regions when the mobility (μh\mu_{\mathrm{h}}) increases, remains constant, or decreases during RuPBA oxidation.

TaskTopology (Input x Hidden Layer x Output)Crossbar Core TypeLearning Rate (L.R.)Other comments
Iris(4 + bias) x 8 x 4Balanced (2 Devices per Syn.)α\alpha=5e-3Constant L.R.
OCR(64 + bias) x 36 x 10Balancedα\alpha=8e-3Constant L.R.
MNIST(784 + bias) x 300 x 10Balancedα\alpha=2.5e-3L.R. schedule (α\alpha/2 epochs 3-5, α\alpha/3 epochs 6-8, α\alpha/4 epochs 9-11, α\alpha/5 after)
f-MNIST(784 + bias) x 300 x 10Balancedα\alpha=1e-3L.R. schedule (same as MNIST)

Table S1. Neural network hyper-parameters used for training

The partial density of states (pDOS) computed by DFT for different oxidation/lithiation states of RuPBA is shown in Figure 6. The ‘fully reduced,’ ‘partially reduced’, and ‘neutral’ forms correspond to compositions of LixRu2[Ru(CN)6]2\mathrm{Li}_{x}\mathrm{Ru}_{2}[\mathrm{Ru}(\mathrm{CN})_{6}]_{2}, where x=4x = 4, 3, and 2, respectively. The partially reduced form (pr-RuPBA), as depicted in Figure 1a, is predicted to have a direct band gap of 0.50 eV, matching the experimentally determined IVCTCP\mathrm{IVCT}_{\mathrm{CP}} maximum for the most conductive state (see Extended Data Table S1). Such agreement between computed band gap and measured IVCT excitation energy was also reported for PB.[71] The fully reduced form is predicted to have a large band gap (4 eV), comparable to the energy of the observed MLCT band (hνmax⁡∼4h\nu_{\max} \sim4 eV). The calculated hole effective masses (mhm_{\mathrm{h}}) for RuPBA range from 0.4 to 0.7 (see Extended Data Table S2); these values are similar to known semiconductors such as p-type germanium.[72] The smallest mhm_{\mathrm{h}} value corresponds the pr-RuPBA state. Based on scattering theory (see DFT Computation for details), we predict a hole charge mobility for pr-RuPBA of 5 cm2^{2}/(V⋅s\mathrm{V}\cdot\mathrm{s}), which is in remarkable agreement with our experimental estimate of 4 cm2^{2}/(V⋅s\mathrm{V}\cdot\mathrm{s}) from the MHT analysis above. The pDOS for pr-RuPBA reveals a delocalized electronic structure, but valence and conduction band states closest to the Fermi energy remain primarily associated with N-Ru sites. This suggests that the lowest energy excitation is between N-Ru sites, which agrees with our interpretation of the IVCT spectra (Figure 5). Thus, both band and electron transfer theories support a mechanism whereby the electrochemically tunable degree of mixed valency between N-Ru centers directly relates to charge carrier concentration.

DFT/HSE06 band structures and density of states of RuPBA in the neutral, partially reduced and fully reduced states

Figure 6. DFT/HSE06 band structures and density of states of RuPBA in the neutral, partially reduced and fully reduced states. The hole effective masses are reported in Extended Data Table 2.

DeviceSmall Digits (OCR) TaskLarge Digits (MNIST) TaskFashion MNIST task
Numeric96.2%98.1%79.4%
RuBPA- (this work)94.9%97.3%78.9%
PEDOT:PSS- Original[1]95.2%96.5%--
PEDOT:PSS- Single[2]95%95.9%79.1%
PEDOT:PSS- Multiple with Centering[2]93.8%92.9%70.3%
LCO[3]96%97.5%--
Bulk-RRAM (TiO2−x_{2-x}, YSZ active layer)[4]--97.5%--
silicon-oxygen-nitrogen-oxygen-silicon (SONOS)[6]--95.8%--

Table S2. Comparison to other NVM device results

We also compute pDOS and mhm_h values for PB at different redox states for comparison (Supplementary Information, § 16). The mhm_h values for PB (Table S4) are larger than those of RuPBA, ranging from 1.3 to 2.9, in agreement with previously published values,[73] which is consistent with the predicted flat bands and localized states (Figure S8), and in agreement with measurements by Long that showed RuPBA is substantially more conductive than PB.[26] The differences between PB and RuPBA band structures primarily arise from their different electron spin configurations (Figure S9). Unlike the high/low spin configuration of PB,[19] both N-coordinated and C-coordinated metal atoms of RuPBA are low spin, allowing for better energy alignment between neighboring Ru atoms and a more delocalized electronic structure (Figure S10–11).

Oxidation StateLattice Constants (A˚\mathrm{\AA}) aaLattice Constants (A˚\mathrm{\AA}) bbLattice Constants (A˚\mathrm{\AA}) ccBand Gap (eV) directBand Gap (eV) indirectmhm_{\mathrm{h}}
neutral7.0817.31710.1822.292.191.3
partially reduced7.2377.28510.2670.360.282.2
fully reduced7.2887.32210.3413.153.152.9

Table S4. DFT/HSE06 optimized lattice parameters, direct and indirect band gaps, and hole effective mass for PB in three different oxidation states

DFT/HSE06 band structures and density of states of PB in the neutral, partially reduced and fully reduced states

Figure S8. DFT/HSE06 band structures and density of states of PB in the neutral, partially reduced and fully reduced states. The hole effective masses are reported in Table S4.

Schematic of the electronic configuration of PB and Ru-PBA in the neutral and fully reduced states

Figure S9. Schematic of the electronic configuration of PB and Ru-PBA in the neutral and fully reduced states.

DFT/HSE06 atom projected DOS for PB and Ru-PBA in the partially reduced state

Figure S10. DFT/HSE06 atom projected DOS for PB and Ru-PBA in the partially reduced state; the atom projections are shown for the 4 metal atoms in the calculation.

Effect of grain boundaries

The estimated mobilities presented above specifically refer to electron-hole transport within the bulk CN-bridged framework. Assuming a hole density (pp) of ca. 102110^{21} cm−3^{-3}, we estimate a bulk crystal conductivity greater than 100 S cm−1^{-1} for the most conductive state (prpr-RuPBA), which is ~10,000 times greater than the measured maximum conductivity. The inkjetprinted films, however, are composed of RuPBA nanocrystallites (Extended Data Figure 8). Measurements of DC conductivity as a function of temperature for as-RuPBA give an activation energy of 0.18 eV (Extended Data Figure 1), which is much higher than the activation energies estimated above from spectroscopy. CT across grain boundaries likely plays a limiting role in the observed film conductivity.[74] Importantly, this activation energy is roughly ¼ of the measured reorganization energy for the IVCTCP\mathrm{IVCT}_{\mathrm{CP}} band (λ=0.64\lambda= 0.64 eV), consistent with very weak coupling and polaron hopping[69, 75] between charge-localized N-Ru centers at the interfaces between neighboring nanocrystals. We propose that electronic coupling is much weaker across grain boundaries due to the absence of NCRuCN bridges between N-coordinated Ru centers.

Outlook

We have shown that the class of active materials for electrochemical transistors extends well beyond extensively studied organic polymers and transition metal chalcogenides into the realm of extended coordination frameworks like RuPBA. Our findings open a window to study electrochemically tunable conductivity in related mixed valence coordination structures such as other PBAs, coordination polymers, and metal-organic frameworks. Two different theoretical approaches, electron transfer theory and Boltzmann transport theory, independently predict similar carrier mobilities within the RuPBA bulk. Our spectroscopic analysis emphasizes that the Marcus-Hush model is not only applicable to CT kinetics in small molecules, but also provides physical insight into mixed valence conductors that belong to the Robin-Day class 2-3 transition regime,[22, 63] exhibiting characteristics of both intervalence hopping and band-like transport due to an intermediate degree of electron delocalization. Both theoretical treatments also support the proposed IVCTCP\mathrm{IVCT}_{\mathrm{CP}} transport pathway. We anticipate major improvements to ECRAM and other electronic applications as devices are made smaller and look forward to taking advantage of RuPBA’s sensitivity to neurotransmitters to achieve communication between biological and artificial neurons.

Methods

All reagents and solvents were purchased from Sigma-Aldrich unless otherwise noted. All water was filtered and deionized to 18.2 MΩ⋅cm\mathrm{M}\Omega\cdot\mathrm{cm} resistivity before use (PureLab flex filtration system, ELGA LabWater). Custom-designed shadow masks for e-beam deposition were purchased from Stencils Unlimited, LLC.

Synthesis of RuPBA

RuPBA was synthesized following the hydrothermal method reported by Long and coworkers.[26] Ruthenium(III) acetylacetonate, Ru(C5H7O2)3\mathrm{Ru(C_5H_7O_2)_3}, (0.500 g, 1.24 mmol) was added to a 20 mL aqueous solution containing 0.260 g (0.310 mmol) of potassium hexacyanoruthenate(II), K4[Ru(CN)6]\mathrm{K_4[Ru(CN)_6]}. The mixture was stirred for 1 hr at room temperature and subsequently transferred to a 45-mL PTFE-lined acid digestion vessel (Parr, model 4744) and heated at 175 °C for 96 hr. The resulting black precipitate was isolated by centrifugation and washed three times with 70 mL of water. To remove Ru(C5H7O2)3\mathrm{Ru(C_5H_7O_2)_3} starting material and potential organic byproducts, the precipitate was rinsed once with 70 mL acetone, three times with 40 mL toluene, and three additional times with 40 mL acetone, using a sonication bath to redisperse the sample in each rinsing solvent and a centrifugation rate of 6000 rpm to collect the precipitate. After rinsing 5 times with 40 mL aliquots of water by sonication and centrifugation (8700 rpm), the final precipitate was redispersed in water and separated into two equal portions. The portion for XRD and elemental analysis was isolated by centrifuge and dried at 60 °C for two days before characterization. Found composition (weight %): K, 6.9; Ru, 45.7; C, 17.6; N, 15.1; H, 1.9. The second portion was isolated by centrifuge and redispersed in DMF to give a 10 mg/mL suspension for preparation of the RuPBA ink.

Preparation of RuPBA ink

Cyclohexanol (2.5 mL) was added to a 5 mL aliquot of 10 mg/mL RuPBA suspension in DMF and sonicated for 10 min. The resulting suspension was centrifuged at 2500 rpm for 5 min to collect and discard large RuPBA aggregates. The remaining suspension, ∼\sim4 mg/mL RuPBA in 2:1 DMF:cyclohexanol (v:v), was used as the RuPBA ink.

Preparation of Li ionogel ink

Lithium bis(fluorosulfonyl)imide, LiFSI, (2.035 g) was dissolved in 10 mL of 1-butyl-1-methylpyrrolidinium bis(trifluoromethanesulfonyl)imide (Pry14_{14}TFSI) ionic liquid by stirring overnight in an Ar-filled glovebox to give a 1.0 M LiFSI in Pry14_{14}TFSI. Preparation of the sol-gel precursor for the ionogel ink was based on the method described by Dunn and coworkers.[40] A 300 μ\muL volume of tetraethylorthosilicate (TEOS) was added to 300 μ\muL of triethoxyvinylsilane (TEVS) and sonicated for 15 min. While under sonication, 300 μ\muL of formic acid was added to the TEOS/TEVS mixture and allowed to sonicate for 15 min to give the sol-gel precursor. A 550 μ\muL aliquot of the sol-gel precursor was then added to 450 μ\muL of the ionic liquid (1.0 M LiFSI in Pry14_{14}TFSI) and sonicated 15 min to give a clear solution, which was subsequently diluted with 1 mL ethanol for use as the Li ionogel ink.

Preparation of protic ionogel ink

A solution containing 17.6 wt% protic ionic liquid (diethylmethylammonium trifluoromethanesulfonate, DEMA:TfO, 98%, IOLITEC), 4.4 wt% poly(vinylidene fluoride-co-hexafluoropropylene), and 78 wt% dimethylacetamide (DMA) was prepared by sonication and subsequently diluted by a factor of 4 with additional DMA solvent.

ECRAM device fabrication

Gold electrode patterns for ECRAM devices and Van der Pauw samples were fabricated on nanosilica-coated polyethylene terephthalate (PET) printing media (Novele™, Novacentrix) via e-beam evaporation using custom-designed shadow masks (Stencils Unlimited) by depositing 5 nm Ti adhesion layer, followed by 150 nm Au top layer. RuPBA and ionogel layers were printed over the Au-patterned PET substrates using a Dimatix Materials Printer (DMP-2850) with DMC-11610 ink cartridges. Removable Scotch tape was used to adhere the backside of the PET substrate to an SiO2_2-coated Si wafer. The fiducial camera of the DMP was used to precisely align the RuPBA prints to the Au electrode pattern array.

RuPBA films for ECRAM devices were printed using a cartridge temperature of 29 °C and substrate temperature of 55 °C. Device channels were printed with one jet nozzle at 20 μ\mum drop spacing for optimal film quality and feature resolution, with a total of 10 print cycles. Channel films included ~75 μ\mum of overlap with the Au contacts. For the device gates, 15 RuPBA print cycles were performed using 11 nozzles at a drop spacing of 31 μ\mum. Two-minute rest periods were inserted in between prints for each device row to keep the cartridge temperature from exceeding 32 ∘C32\,^\circ\mathrm{C}, which was found to help prevent nozzle clogging. The resulting RuPBA films were then dried at 60 ∘C60\,^\circ\mathrm{C} for at least 48 hours. After realignment of the substrate, ionogel films were printed at room temperature with 20 μm20\,\mu\mathrm{m} drop spacing using 3 jetting nozzles and at least 6 print layers to ensure full coverage.

ECRAM devices for dopamine sensors included an additional 4 by 4 mm area to the gate to serve as the base of the reservoir for the dopamine solution. For these devices, RuPBA was printed over the entire area of Au gate electrode and the channel region, similar to the other ECRAM devices. The ionogel was printed over the region shown in Figure 3a, leaving a 4 by 4 mm RuPBA/Au region without ionogel coating. An acrylic adhesive backed plastic well was fabricated by laser cutting and attached to the device. The well was initially filled with 25 μ\muL of 0.1 M sodium phosphate buffer, pH 5.5. A pipet was used to incrementally add 1 μ\muL aliquots of the dopamine solution (0.1 mM dopamine hydrochloride, 0.1 M sodium phosphate buffer, pH 5.5) to the reservoir.

X-ray diffraction

Powder x-ray diffraction measurements were performed on a PANalytical Empyrean diffractometer equipped with a PIXcel3D detector and operated at 44 kV and 40 kA using Cu KαK\alpha radiation (λ=1.5406\lambda= 1.5406 Å). A reflection-transmission spinner was used as a sample holder and the spinning rate was set at 4 rpm. The patterns were collected in the 2θ2\theta range of 5 to 75°, and the step size was 0.026°.

X-ray photoelectron spectroscopy

Samples for x-ray photoelectron spectroscopy were prepared by dropcasting the RuPBA ink on a Au-coated SiO₂/Si substrate. The spectra were acquired using Al KαK\alpha: 1486.6 eV source and monochromator (Scientа Omicron, Inc., XM 1000Mkii) under ultrahigh vacuum.

RuPBA film conductivity and ECRAM pulsed-programming measurements

Electrical measurements of Van der Pauw samples and pulsed conductance tuning of ECRAM devices were performed under N₂ atmosphere at slightly reduced pressure (~600 torr) and controlled temperature using a CPX-VF probe station and temperature controller (Lakeshore).

Sheet resistance measurements were performed with a Keysight B1500A semiconductor device analyzer to obtain conductivity based on the measured film thickness. Current-voltage recordings were also performed at each temperature to make sure the response remained ohmic.

Pulsed programming and state retention testing of ECRAM devices was performed using a custom LabVIEW waveform generation program, as previously described,[32, 33, 76] to control analog voltage outputs and record input current to an NI-DAQ PCIe-6363 at 10 kHz sampling. To switch the ECRAM device gate to open circuit during the ‘read’ steps, a CMOS analog switch (MAX327, MAXIM Integrated Products) was implemented and powered by a dual output DC power supply (E3620A, Keysight). A low-noise current amplifier (FEMTO, DLPCA-200) was used to measure the current at 10610^{6} V/A gain and 200 kHz low-pass analog filter. The gate charge, ΔQ\Delta Q, resulting from each write pulse was measured by integration of the raw current-time traces using the cumulative trapezoidal numerical integration function in MATLAB 2019.

Cell Culture

Neuro2a cells were purchased from ATCC (CCL-131) and cultured in high-glucose DMEM (Fisher Scientific #11995073) with 10% Fetal Bovine Serum (FBS) (Seradigm Premium Grade, Avantor) and 1% Penicillin/Streptomycin (Thermo Fisher #15140122). Cells were grown at 37 °C and 5% CO₂ and passaged/harvested using Trypsin-EDTA (Fisher Scientific #25300120).

Growth and Viability Assays

A 1 mg/mL aqueous dispersion of RuPBA was spotted onto wells of a 24-well plate (200 ul per well) and allowed to dry at room temperature for 1 hour. Cells were seeded at densities of 10,000, 25,000 and 50,000 cells/well in quadruplicate, and allowed to grow for 3 days, either in the presence or absence of RuPBA. At the end of 3 days, cells were harvested, stained with Trypan Blue (1:1 ratio, incubated for 5 minutes at room temperature) and counted using the BioRad TC20 Cell Counter. Pictures of the cells on the RuPBA substrate were taken using brightfield imaging with the EVOS M5000 imaging system (at 200X magnification).

Scanning Electron Microscopy

Cells were prepared for scanning electron microscopy as previously described.[77] Briefly, cells were plated on Nunc Thermanoх coverslips (Thermo Fisher #150067) and grown for 3 days in DMEM media with 10% FBS (see above). The coverslips were then fixed in 2.5% glutaraldehyde (Fisher Scientific #O2957-1) at room temperature for 30 minutes, washed 3 x 2 minutes with rinsing buffer (phosphate buffer), and post-fixed with 1% osmium tetroxide (Sigma #20816-12-0). After another 2-minute wash with rinsing buffer and 2 x 2-minute washes with deionized water, the coverslips were dried with anhydrous ethanol. Samples were then dehydrated using a CO2CO_2 critical point dryer, sputter coated with ∼1\sim1 nm Au, and imaged using a FEI Nova NanoSEM 450 at 5 kV accelerating voltage.

In-situ UV/Vis/NIR and Raman spectroscopy

All spectroelectrochemical measurements were performed using a BioLogic SP-300 bipotentiostat.

In-situ UV/Vis/NIR spectroscopic interrogation of RuPBA films on ITO-coated glass slides (Sigma Aldrich) was performed using a JASCO V-770 spectrophotometer equipped with an integrating sphere. The spectroelectrochemical cell consisted of inkjet-printed RuPBA on ITO-coated glass (working electrode), a gold wire (counter electrode), a Ag/AgCl wire (quasireference electrode), and a quartz window in a custom-designed Teflon cell with rubber gaskets to seal electrolyte. The electrolyte contained 1.0 M LiFSI dissolved in Pry14_{14}TFSI.

The electrochromic cell shown in the Video S1 of the Supplementary Information was constructed using two FTO-coated glass electrodes, each with inkjet-printed RuPBA films, and double-sided Kapton tape to seal the 1.0 M LiFSI in Pry14_{14}TFSI electrolyte between the two electrodes. The working electrode consisted of square shaped RuPBA printed film of 2.25 cm² area. The printed RuPBA area of the counter/reference electrode was a square frame with inner and outer square edge lengths of 3 and 4 cm (7 cm² frame area).

In-situ NIR absorption spectra of ECRAM devices were acquired with an NIRQuest+2.5 spectrometer from Ocean Insight. Fiber optic cables for 400-2400 nm wavelength range (Thorlabs, AFS50/125Y) were used with collimating lenses to direct the beam path of the tungsten-halogen light source through the ECRAM device using an adjustable stage for film transmission measurements (Stage-RTL-T, Ocean Insight). Blank reference measurements were performed by positioning the light beam (∼0.2\sim0.2 mm cross-sectional diameter) at an area of the ECRAM device that included the glass microscope slide, PET substrate, and ionogel, but not RuPBA. The sample was then repositioned to pass the incident light through the middle of the RuPBA channel. Constant potentials were applied to the gate for 725 s using a BioLogic SP-300 bipotentiostat. Afterwards, the gate was set to open circuit and the channel conductance was monitored while simultaneously recording spectra (0.5 s integration time, 10 scans averaged for each spectrum).

In-situ Raman spectra were acquired with an inVia{sup}TM Raman microscope (Renishaw) using a 633 nm laser at 0.1% power. Using a 50X objective, the beam was focused to an area where the RuPBA channel film coats one of the Au contacts. The plasmonic enhancement from Au was found to be necessary to achieve sufficient signal. Acquisition settings: 50 s exposure time, 60 accumulations (3000 s total acquisition time per spectrum).

DFT Computation

All density functional theory calculations were performed using the Vienna ab initio simulation package (VASP).[78] The geometry and electronic properties were predicted using the hybrid HSE06 exchange-correlation functional within the spin polarization formulism. The planewave basis set energy cut-off was set to 500 eV and a 3×3×23 \times3 \times2 k-point grid was using during geometry optimization (atomic positions and cell volume). The geometries were optimized until all forces were below 0.02 eV/Å. A single-point calculation was then preformed, with a k-point grid of 6×6×46 \times6 \times4, to determine the density of states. The band structures and hole effective masses (mhm_{\mathrm{h}}) were determined using AMSET[79] which depends on BoltzTrap2[80] for interpolation.

Seekpath[81] was used to determine the high symmetry band paths; note, some paths have been omitted in Figure 6 and Figure S9 for aesthetics reasons because they looked visually similar. The lowest mhm_{\mathrm{h}} determine via AMSET at gamma are reported. Three oxidization states were considered for both PB and RuPBA; fully reduced (fr), partially reduced (pr), and neutral. The different oxidization states were achieved by varying the number of Li atoms (4, 3 and 2 Li atoms respectively; note, PB calculations were performed with K in place of Li). An orthorhombic unit cell was used containing 4 Ru (see Figure S7 of Supplementary Information) which is the smallest system allowing for a pr state to be represented.

DFT/HSE06 optimized geometry of partially reduced RuPBA

Figure S7. DFT/HSE06 optimized geometry of partially reduced RuPBA, Li3Ru2[Ru(CN)6]2\mathrm{Li}_{3}\mathrm{Ru}_{2}[\mathrm{Ru}(\mathrm{CN})_{6}]_{2}. Lattice parameters are reported in Extended Data Table S2.

The hole charge mobility was calculated for pr-RuPBA with the PBEsol+U functional using a Hubbard UeffU_{\mathrm{eff}} value of 4.0 eV (resulting band gap is in good agreement with HSE06). The planewave basis set energy cut-off was set to 500 eV and a 4×4×34 \times4 \times3 k-point grid was used during geometry optimization; the geometry was relaxed until all forces were less 0.0005 eV/A˚0.0005\ \mathrm{eV}/\mathring{\mathrm{A}}. The elastic constants, effective phonon frequency, static and high-frequency dielectric constants and deformation potential were calculated at the same level on theory. A signal-point calculation using a denser k-point mesh, 8×8×68 \times8 \times6, was used to obtain the wave function coefficients and for band structure interpolation. These properties/calculations were used to determine acoustic deformation, ionized impurity, and polar optical phonon scattering. The reported hole mobility is calculated for a carrier concentration of 1.8×1021 cm−31.8 \times10^{21}\ \mathrm{cm}^{-3} which corresponds to 1 carrier per unit volume. The charge mobility calculation was performed using AMSET[79].

Extended Data

Full conductance trace showing one million write pulses and more than 6000 potentiation/depression ramps

Figure Extended Data Fig. 4. Full conductance trace resulting from one million write pulses and more than 6000 potentiation/depression ramps for a RuPBA EC-RAM device with Li ionogel electrolyte.

EE / Vλ\lambda / eVhΔν1/2h\Delta\nu_{1/2} / eVHeffH_{\mathrm{eff}} / eVEa∗E_{\mathrm{a}}^{*} / meVνel\nu_{\mathrm{el}} / 101410^{14} s−1^{-1}kctk_{\mathrm{ct}} / 101210^{12} s−1^{-1}μh\mu_{\mathrm{h}} / cm2^{2} V−1^{-1} s−1^{-1}
0.000.810.640.20507.68.90.95
0.050.760.500.252412252.7
0.100.740.450.261613343.7
0.150.660.390.241112414.3
0.300.590.360.221110424.5
0.600.510.340.18117.5414.3
1.000.550.390.18177.3333.5

Table 1. Measured parameters from deconvolution of IVCTCP_{\mathrm{CP}} bands at varied applied potentials, as determined from the spectra shown in Figure 2a. Parameters are defined in the main text and § 15 of the Supplementary Information.

Oxidation StateLattice Constants (Å)Band Gap (eV)mhm_h
abcdirectindirect
neutral7.2697.29310.3341.260.980.70
partially reduced7.4157.21710.2950.500.100.41
fully reduced7.5437.03810.1804.294.290.52

Table 2. DFT/HSE06 optimized lattice parameters, direct and indirect band gaps, and hole effective mass for RuPBA in three different oxidation states.

Acknowledgements

This research was supported by the Laboratory-Directed Research and Development (LDRD) Program. AAT and MJM were also supported by the DOE Office of Science Research Program for Microelectronics Codesign (sponsored by ASCR, BES, HEP, NP, and FES) through the Abisko Project, PM Robinson Pino (ASCR). Sandia National Laboratories is a multimission laboratory managed and operated by National Technology and Engineering Solutions of Sandia, LLC., a wholly owned subsidiary of Honeywell International, Inc., for the U.S. Department of Energy’s National Nuclear Security Administration under Contract DE-NA-0003525. The views expressed in this article do not necessarily represent the views of the U.S. Department of Energy or the United States Government.

Charge analysis of RuPBA ECRAM transfer characteristics

ECRAM device architecture and fabrication

ECRAM device fabrication by inkjet printing, showing device architecture and fabrication flow, device array on transparent PET substrate, and an individual printed device

Figure S2. ECRAM device fabrication by inkjet printing. a, Device architecture and flow chart of fabrication process. b, Photo of device array on transparent PET substrate. c, Photo of individual printed device.

Simulation of ECRAM synapse array for online learning accuracy predictions

We evaluated the implications for online learning by simulating a hardware multi-layer perceptron, physically constructed of crossbars of various sizes using the CrossSim codebase (Table S1) populated with simulated RuBPA synapses learning with stochastic gradient descent (batch size 1). Synaptic switching linearity and symmetry was based on look up tables (Figure S3) generated from the experimental data (Extended Data Figure S4).

Maps of RuPBA ECRAM switching magnitude and probability during potentiation and depression

Figure S3. Maps of RuPBA ECRAM switching magnitude and probability as a function of conductance (look up tables) during a, potentiation and b, depression, originating from recording shown in Extended Data Figure S4. Median probability is the yellow/white region along the middle.

Our results are comparable to that of other leading linear and symmetric nonvolatile memory (NVM) candidates, as show comprehensively in Table S2. RuPBA devices achieves parity with or slightly outperforms synapses based on PEDOT/PSS, [1, 2] lithium cobalt oxide (LCO),[3] and bulk-RRAM devices based on TiO2−x_{2-x}.[4] While our simulations do not incorporate realistic device-to-device effects as in Li et al.,[2] additional agreement with numeric training can be achieved by centering synapses to a tighter, shared workable range. As this range is the most linear and symmetric range regardless of RuBPA geometry, this augurs well for the future of these devices in online learning and inference contexts. Along with a variety of promising NVM devices with a degree of high linearity and symmetry,[4, 5] these results establish RuBPA ECRAM devices as a leading candidate for future embedded/edge systems that require online learning capabilities.

Analysis of switching time and energy for ECRAM synapse scaling relationship

To evaluate the upper limit to the device programming rate, one must first determine a threshold criterion for the minimum conductance change, ΔGmin⁡\Delta G_{\min}, that can be accurately measured. The number of discrete programmable conductance states within a specified range ultimately depends on the accuracy of the measured conductance, which is dictated by the total noise level of the read current. The standard deviation from all measured ΔG\Delta G values for multiple potentiation-depression pulsed programming cycles (σΔG\sigma_{\Delta G}) was selected to set the threshold criterion as ΔG≥2σΔG\Delta G \ge2\sigma_{\Delta G} with the limiting case, ΔGmin⁡=2σΔG\Delta G_{\min}=2\sigma_{\Delta G}. For a fair comparison, the applied write pulse durations (Δtw\Delta t_{\mathrm{w}}) were varied for each device size to achieve ∼\sim70 pulses per ramp, giving roughly the same average change in conductance per switching pulse (ΔGavg\Delta G_{\mathrm{avg}}). The average σΔG\sigma_{\Delta G} for the three different device sizes was 0.5(±\pm0.1) nS, giving ΔGmin⁡≈1\Delta G_{\min}\approx1 nS (See Table S1). To evaluate switching speed, we estimate the minimum write pulse duration, Δtmin⁡\Delta t_{\min}, based on the average programming rate, ΔGavg/Δtw\Delta G_{\mathrm{avg}}/\Delta t_{\mathrm{w}}, to give ΔGavg/Δtw=ΔGmin⁡/Δtmin⁡\Delta G_{\mathrm{avg}}/\Delta t_{\mathrm{w}}=\Delta G_{\min}/\Delta t_{\min}, which leads to

Δtmin⁡=ΔGmin⁡ΔGavg×Δtw(3)\Delta t_{\min}=\frac{\Delta G_{\min}}{\Delta G_{\mathrm{avg}}}\times\Delta t_{\mathrm{w}} \tag*{(3)}

The minimum energy requirement for each weight update, ΔEmin⁡\Delta E_{\min}, is the product of the voltage applied to the gate, Vg=1V_{\mathrm{g}}=1 V, and the amount of transferred charged, ΔQmin⁡\Delta Q_{\min}, to achieve the minimum conductance change (ΔEmin⁡=VgΔQmin⁡\Delta E_{\min}=V_{\mathrm{g}}\Delta Q_{\min}). To determine ΔQmin⁡\Delta Q_{\min}, we first measured each increment of transferred charge, ΔQ\Delta Q, by integration of the current-time response for each write pulse (eq 3).

ΔQ=∫t1t2iw dt(4)\Delta Q=\int_{t_1}^{t_2}i_{\mathrm{w}}\,\mathrm{d}t \tag*{(4)}

We then use the ΔGmin⁡/ΔGavg\Delta G_{\min}/\Delta G_{\mathrm{avg}} ratio to approximate ΔQmin⁡\Delta Q_{\min} from the measured average charge, ΔQavg\Delta Q_{\mathrm{avg}}.

ΔQmin⁡≅ΔGmin⁡ΔGavg×ΔQavg(5)\Delta Q_{\min}\cong\frac{\Delta G_{\min}}{\Delta G_{\mathrm{avg}}}\times\Delta Q_{\mathrm{avg}} \tag*{(5)}

Detailed description of dopamine sensor operation

Figure S4 shows the simultaneously recorded gate voltage and current that accompanies the source-drain conductance trace in Figure 3d of the main text for dopamine sensing measurements. During the first 590 s, the device reservoir contains phosphate buffer without dopamine. The first applied gate voltage pulse (p1p_1) switches the conductance to a lower value and the second pulse (p2p_2) returns the conductance to the initial value (G0=4.85 μSG_0 = 4.85\,\mu\mathrm{S}). The open-circuit voltage between the gate and channel (VgV_g) responds similarly, switching to a more positive value after p1p_1 and returning to the initial value (−80 mV-80\,\mathrm{mV}) after p2p_2. Upon introduction of 4 μM4\,\mu\mathrm{M} DA (step 1), VgV_g shifts to more negative potential as the dopamine undergoes redox with the RuPBA at the gate because the decrease of mixed valence II:III sites causes the Fermi level of the gate to increase (Figure S4). Meanwhile, the channel conductance remains at 4.85 μS4.85\,\mu\mathrm{S} because the gate is still at open circuit. When the next p1p_1 pulse is applied (step 2), the gate oxidation current jumps to a higher magnitude than in the previous three p1p_1 pulses because the dopamine reaction generated RuII^{\mathrm{II}} sites to (re)oxidize, thus lowering the Fermi level of the gate and causing the opposite reaction (reduction) at the channel. After the application of p2p_2, the gate-channel circuit is switched open (step 3), and the resulting conductance is lower than the previous value when no dopamine was present. The net result of steps 1-3 is to increase RuII^{\mathrm{II}} sites and decrease conductivity at all regions of the device, starting with the gate (step 1) and ending with the channel (steps 2-3).

IVCT spectral analysis based on electron transfer theory

Spectroscopic determination of the reorganization energy, electronic coupling, and activation energy followed the methods detailed by Brunschwig, Creutz, and Sutin based on a two-state model for a class 2B system.[7] The IVCT band manifold (νmax⁡=Eop/h\nu_{\max}=E_{\mathrm{op}}/h) represents the combination of reorganization energy (λ\lambda) and the energy difference between the reactant and product ground states (ΔE0\Delta E_0). For IVCTCP_{\mathrm{CP}}, ΔE0=0\Delta E_0=0 because the reactant and product states are essentially identical, leading to eq (6) to determine λ\lambda.

hνmax⁡=Eop≈λ(6)h\nu_{\max}=E_{\mathrm{op}}\approx\lambda \tag*{(6)}

For such symmetrical systems, one can approximate the electronic coupling (HeffH_{\mathrm{eff}}) based on the peak shape following a two-state model (eq (7))

Δν1/2=[4ln⁡(2)λkBT]1/2+λ−2Heff(7)\Delta\nu_{1/2}=\left[4\ln(2)\lambda k_{\mathrm{B}}T\right]^{1/2}+\lambda-2H_{\mathrm{eff}} \tag*{(7)}

where Δν1/2\Delta\nu_{1/2} is the full width at half maximum for the fitted IVCTCP_{\mathrm{CP}} band. The activation energy, Ea∗E_{\mathrm{a}}^{*}, was calculated by eq (8).

Ea∗=(λ−2Heff)24λ(8)E_{\mathrm{a}}^{*}=\frac{(\lambda-2H_{\mathrm{eff}})^2}{4\lambda} \tag*{(8)}

This equation modifies the more well-known expression for self-exchange reactions[8] (Ea=λ/4E_{\mathrm{a}}=\lambda/4) by including the influence of electronic coupling on the activation energy. For a Robin–Day class 2 system, λ>2Heff\lambda>2H_{\mathrm{eff}} and Ea∗>0E_{\mathrm{a}}^{*}>0. For class 3, λ<2Heff\lambda<2H_{\mathrm{eff}} and the activation barrier does not exist.

The electronic transmission coefficient, κ\kappa, describes the probability of electron transfer when the energy barrier is minimized according to the Landau–Zener model,[8, 9] as expressed by eq (9),

κ=2−2exp⁡(−νel2νn)2−exp⁡(−νel2νn)(9)\kappa=\frac{2-2\exp\left(-\frac{\nu_{\mathrm{el}}}{2\nu_n}\right)}{2-\exp\left(-\frac{\nu_{\mathrm{el}}}{2\nu_n}\right)} \tag*{(9)}

where νel\nu_{\mathrm{el}} is the electronic frequency factor, calculated by eq (10).[8]

νel=2πHeff2ℏ4πλkBT(10)\nu_{\mathrm{el}}=\frac{2\pi H_{\mathrm{eff}}^2}{\hbar\sqrt{4\pi\lambda k_{\mathrm{B}}T}} \tag*{(10)}

In spectroelectrochemical measurements (Figure 4a of the main text), the spectrum of fully reduced RuPBA was used as a reference to subtract from spectra of the partially oxidized states, giving Figure S5. This was done to isolate the IVCT bands from the transitions that do not change during oxidation. This corrected absorbance was divided by wavenumber (Abs/ν\nu) to give the ‘reduced’ spectra,[10] as shown in Figure S5, which provides a more accurate determination of λ\lambda and HeffH_{\mathrm{eff}}, [7] and the bands were fit as asymmetric gaussian peaks. Example deconvolutions and graphically represented measurements of the IVCTCP_{\mathrm{CP}} band parameters are shown in Figure S6. The IVCTCP_{\mathrm{CP}} peaks for the reduced spectra were more ambiguous at higher potentials because νmax⁡\nu_{\max} shifts to a lower energy than the experimental cutoff (Figure S6). To best approximate the partial peak width at the lower energy side, the fitted peak was constrained such that Abs/ν\nu approaches zero at ν=0\nu= 0. This approach was guided by the observation of decreasing symmetry and broadness of the unfitted spectra at the lower potential range (ca. -0.1 to 0.15 V, as shown in the inset of Figure S5). Similar to the IVCTCP_{\mathrm{CP}} fit, the best fit for IVCTNN_{\mathrm{NN}} became more asymmetric at more oxidizing potentials as the coupling increases. The measured values of λ\lambda and Δν1/2\Delta\nu_{1/2} were used to determine HeffH_{\mathrm{eff}} and Ea∗E_{\mathrm{a}}^{*}, as listed in Extended Data Table S1. A similar method was followed for the spectrum of as-synthesized RuPBA, as shown in Figure 5a of the main text.

Detailed analysis of absorption spectrum evolution during RuPBA oxidation

Figure S5. a, Spectra from Figure 2a in the main text after subtracting the spectrum of frfr-RuPBA at -0.2 V. Vertical red lines mark the absorbance maxima at different oxidation states. b, Measured wavenumber for the band manifold as a function of applied potential. c, ‘Reduced’ spectra from part a above after dividing by ν\nu.

Deconvolution of IVCT bands for reduced absorption spectra at $E = 0$ V and $E = 1.0$ V

Figure S6. Deconvolution of IVCT bands for reduced absorption spectra collected at a, E=0E = 0 V and b, E=1.0E = 1.0 V. These examples were taken from the overlaid spectra shown in Figure 2a of the main text after subtracting the absorbance spectrum at E=−0.2E = -0.2 V and dividing the absorbance by wavenumber.

Predictions from DFT

DFT/HSE06 charge density difference for adding a hole to pr-RuPBA

Figure S11. DFT/HSE06 charge density difference for adding a hole (removing an electron) to pr-RuPBA. The figure illustrates that the hole delocalizes; moreover, the majority of the density is on the two RuN sites in the cell but some is on a RuC site. The delocalization, in the NN direction, is believed to be why the effective mass is lower for RuPBA than PB.

Optical profilometry and roughness mapping of inkjet-printed RuPBA film

Optical image, heat map, and 3-D profile of a RuPBA channel

Figure S12. Optical profiles of RuPBA channel for an EC-RAM device without electrolyte. a, Optical image showing selected region of the channel to map. b, Heat map and c, corresponding 3-D profile. Blue region indicates the PET substrate. Measured film thickness is 0.6 ±\pm 0.1  μm\ \mu\mathrm{m} based on 6 measurements at different edge regions of the channel film.

X-ray photoelectron spectroscopy

Figure S13 shows the photoelectron spectra for as-synthesized RuPBA. The maximum intensity for Ru 3d5/2\mathrm{Ru}\ 3d_{5/2} occurs at a binding energy (BE) of 280.5 eV, which is close to that reported for K4RuII(CN)6\mathrm{K}_{4}\mathrm{Ru}^{\mathrm{II}}(\mathrm{CN})_{6} (280.9 eV),[11] suggesting a mostly reduced state, consistent with the high potassium content from elemental analysis (see Methods) and spectroscopic observations discussed in the main text. The Ru regions were best fit with two peaks of BE separated by ∼1\sim1 eV (red and blue peak fits). The components at higher BE (blue fits) may at least partially correspond to RuIII\mathrm{Ru}^{\mathrm{III}} content, but the relative intensity seems too high to assign to charge-localized RuIII\mathrm{Ru}^{\mathrm{III}} sites considering the high degree of electronic delocalization in RuPBA. Also, the BE difference is half the expected value of 2 eV for comparing RuII\mathrm{Ru}^{\mathrm{II}} to RuIII\mathrm{Ru}^{\mathrm{III}} in complexes with identical composition.[12] We therefore tentatively assign the two components to different coordination with either the N or C atoms of the cyanide bridging ligand, in addition to H2O\mathrm{H}_{2}\mathrm{O} bound at defect sites.

Heat map and 3-D profile of a RuPBA Van der Pauw film

Figure S13. Optical profiles of RuPBA Van der Pauw film for conductivity measurements. a, Example of heat map and b, corresponding 3-D profile focused at a section showing the edge of the film. Blue region indicates the PET substrate. Measured film thickness is 0.6 ±\pm 0.1  μm\ \mu\mathrm{m} from 15 measurements at different edge regions of the channel film.

X-ray photoelectron spectra of as-synthesized RuPBA showing Ru 3p, Ru 3d, N 1s, and O 1s regions

Figure S14. X-ray photoelectron spectra of as-synthesized RuPBA showing a, Ru 3p, b, Ru 3d, c, N 1s, and d, O 1s regions.

References

  1. [1]Sood A, Poletayev AD, Cogswell DA, Csernica PM, Mefford JT, Fraggedakis D, et al. Electrochemical ion insertion from the atomic to the device scale. Nature Reviews Materials 2021, 6(9): 847-867.DOI
  2. [2]Keggin JF, Miles FD. Structures and Formulae of the Prussian Blues and Related Compounds. Nature 1936, 137(3466): 577-578.DOI
  3. [3]Simonov A, De Baerdemaeker T, Boström HLB, Ríos Gómez ML, Gray HJ, Chernyshov D, et al. Hidden diversity of vacancy networks in Prussian blue analogues. Nature 2020, 578(7794): 256-260.DOI
  4. [4]Avila Y, Acevedo-Peña P, Reguera L, Reguera E. Recent progress in transition metal hexacyanometallates: From structure to properties and functionality. Coord Chem Rev 2022, 453: 214274.DOI
  5. [5]Song X, Song S, Wang D, Zhang H. Prussian Blue Analogs and Their Derived Nanomaterials for Electrochemical Energy Storage and Electrocatalysis. Small Methods 2021, 5(4): 2001000.DOI
  6. [6]Yi H, Qin R, Ding S, Wang Y, Li S, Zhao Q, et al. Structure and Properties of Prussian Blue Analogues in Energy Storage and Conversion Applications. Adv Funct Mater 2021, 31(6): 2006970.DOI
  7. [7]Zhou A, Cheng W, Wang W, Zhao Q, Xie J, Zhang W, et al. Hexacyanoferrate-Type Prussian Blue Analogs: Principles and Advances Toward High-Performance Sodium and Potassium Ion Batteries. Advanced Energy Materials 2021, 11(2): 2000943.DOI
  8. [8]Wu X, Hong JJ, Shin W, Ma L, Liu T, Bi X, et al. Diffusion-free Grotthuss topochemistry for high-rate and long-life proton batteries. Nat Energy 2019, 4(2): 123-130.DOI
  9. [9]Ono K, Ishizaki M, Kanaizuka K, Togashi T, Yamada T, Kitagawa H, et al. Grain-Boundary-Free Super-Proton Conduction of a Solution-Processed Prussian-Blue Nanoparticle Film. Angew Chem Int Ed 2017, 56(20): 5531-5535.DOI
  10. [10]Pearce J. Studies of any toxicological effects of Prussian blue compounds in mammals—A review. Food Chem Toxicol 1994, 32(6): 577-582.pubmed.ncbi.nlm.nih.gov/8045465
  11. [11]Kong B, Selomulya C, Zheng G, Zhao D. New faces of porous Prussian blue: interfacial assembly of integrated hetero-structures for sensing applications. Chem Soc Rev 2015, 44(22): 7997-8018.DOI
  12. [12]Cobo S, Molnár G, Carcenac F, Szilágyi PA, Salmon L, Vieu C, et al. Thin films of Prussian blue: sequential assembly, patterning and electron transport properties at the nanometric scale. J Nanosci Nanotechnol 2010, 10(8): 5042-5050.pubmed.ncbi.nlm.nih.gov/21125848
  13. [13]Song YY, Jia W-Z, Li Y, Xia X-H, Wang Q-J, Zhao J-W, et al. Synthesis and Patterning of Prussian Blue Nanostructures on Silicon Wafer via Galvanic Displacement Reaction. Adv Funct Mater 2007, 17(15): 2808-2814.DOI
  14. [14]Manumpil MA, Leal-Cervantes C, Hudson MR, Brown CM, Karunadasa HI. Electronic Conductivity in a Porous Vanadyl Prussian Blue Analogue upon Air Exposure. Inorg Chem 2017, 56(21): 12682-12686.DOI
  15. [15]Hosseini P, Wolkersdörfer K, Wark M, Redel E, Baumgart H, Wittstock G. Morphology and Conductivity of Copper Hexacyanoferrate Films. The Journal of Physical Chemistry C 2020, 124(31): 16849-16859.DOI
  16. [16]Tennakone K, Dharmaratne WGD. Experimental and theoretical study of electronic conduction in H2O-doped Prussian blue. Journal of Physics C: Solid State Physics 1983, 16(29): 5633-5639.DOI
  17. [17]Rosseinsky DR, Tonge JS, Berthelot J, Cassidy JF. Site-transfer conductivity in solid iron hexacyanoferrates by dielectric relaxometry, voltammetry and spectroscopy. Prussian Blue, congeners and mixtures. Journal of the Chemical Society, Faraday Transactions 1: Physical Chemistry in Condensed Phases 1987, 83(2): 231-243.DOI
  18. [18]Xidis A, Neff VD. On the Electronic Conduction in Dry Thin Films of Prussian Blue, Prussian Yellow, and Everitt's Salt. J Electrochem Soc 1991, 138(12): 3637-3642.DOI
  19. [19]Pajerowski DM, Watanabe T, Yamamoto T, Einaga Y. Electronic conductivity in Berlin green and Prussian blue. Physical Review B 2011, 83(15): 153202.DOI
  20. [20]Piechota EJ, Meyer GJ. Introduction to Electron Transfer: Theoretical Foundations and Pedagogical Examples. J Chem Educ 2019, 96(11): 2450-2466.DOI
  21. [21]Demadis KD, Hartshorn CM, Meyer TJ. The Localized-to-Delocalized Transition in Mixed-Valence Chemistry. Chem Rev 2001, 101(9): 2655-2686.pubmed.ncbi.nlm.nih.gov/11749392
  22. [22]Brunschwig BS, Creutz C, Sutin N. Optical transitions of symmetrical mixed-valence systems in the Class II–III transition regime. Chem Soc Rev 2002, 31(3): 168-184.pubmed.ncbi.nlm.nih.gov/12122642
  23. [23]D'Alessandro DM, Keene FR. Intervalence Charge Transfer (IVCT) in Trinuclear and Tetranuclear Complexes of Iron, Ruthenium, and Osmium. Chem Rev 2006, 106(6): 2270-2298.
  24. [24]Murase R, Leong CF, D’Alessandro DM. Mixed Valency as a Strategy for Achieving Charge Delocalization in Semiconducting and Conducting Framework Materials. Inorg Chem 2017, 56(23): 14373-14382.DOI
  25. [25]Shatruk M, Dragulescu-Andrasi A, Chambers KE, Stoian SA, Boinnaar EL, Achim C, et al. Properties of Prussian Blue Materials Manifested in Molecular Complexes: Observation of Cyanide Linkage Isomerism and Spin-Crossover Behavior in Pentanuclear Cyanide Clusters. Journal of the American Chemical Society 2007, 129(19): 6104-6116.pubmed.ncbi.nlm.nih.gov/17455931
  26. [26]Behera JN, D’Alessandro DM, Soheilnia N, Long JR. Synthesis and Characterization of Ruthenium and Iron–Ruthenium Prussian Blue Analogues. Chem Mater 2009, 21(9): 1922-1926.DOI
  27. [27]Talin AA, Centrone A, Ford AC, Foster ME, Stavila V, Haney P, et al. Tunable Electrical Conductivity in Metal-Organic Framework Thin-Film Devices. Science 2014, 343(6166): 66-69.DOI
  28. [28]Aubrey ML, Wiers BM, Andrews SC, Sakurai T, Reyes-Lillo SE, Hamed SM, et al. Electron delocalization and charge mobility as a function of reduction in a metal–organic framework. Nat Mater 2018, 17(7): 625-632.DOI
  29. [29]van de Burgt Y, Melianas A, Keene ST, Malliaras G, Salleo A. Organic electronics for neuromorphic computing. Nature Electronics 2018, 1(7): 386-397.
  30. [30]Rivnay J, Inal S, Salleo A, Owens RM, Berggren M, Malliaras GG. Organic electrochemical transistors. Nature Reviews Materials 2018, 3(2): 17086.
  31. [31]van de Burgt Y, Lubberman E, Fuller EJ, Keene ST, Faria GC, Agarwal S, et al. A non-volatile organic electrochemical device as a low-voltage artificial synapse for neuromorphic computing. Nat Mater 2017, 16(4): 414-418.DOI
  32. [32]Fuller EJ, Keene ST, Melianas A, Wang Z, Agarwal S, Li Y, et al. Parallel programming of an ionic floating-gate memory array for scalable neuromorphic computing. Science 2019, 364(6440): 570-574.DOI
  33. [33]Li Y, Xiao TP, Bennett CH, Isele E, Melianas A, Tao H, et al. In situ Parallel Training of Analog Neural Network Using Electrochemical Random-Access Memory. Front Neurosci 2021, 15: 636127-636127.DOI
  34. [34]Keene ST, Lubrano C, Kazemzadeh S, Melianas A, Tuchman Y, Polino G, et al. A biohybrid synapse with neurotransmitter-mediated plasticity. Nat Mater 2020, 19(9): 969-973.DOI
  35. [35]Marcus RA. On the Theory of Oxidation-Reduction Reactions Involving Electron Transfer. I. The Journal of Chemical Physics 1956, 24(5): 966-978.
  36. [36]Hush NS. Adiabatic theory of outer sphere electron-transfer reactions in solution. Transactions of the Faraday Society 1961, 57(0): 557-580.DOI
  37. [37]Fuller EJ, Gabaly FE, Léonard F, Agarwal S, Plimpton SJ, Jacobs-Gedrim RB, et al. Li-Ion Synaptic Transistor for Low Power Analog Computing. Adv Mater 2017, 29(4): 1604310.DOI
  38. [38]Tang J, Bishop D, Kim S, Copel M, Gokmen T, Todorov T, et al. ECRAM as Scalable Synaptic Cell for High-Speed, Low-Power Neuromorphic Computing. 2018 IEEE International Electron Devices Meeting (IEDM); 2018 1-5 Dec. 2018; 2018. p. 13.1.1-13.1.14.DOI
  39. [39]Ishizaki M, Ando H, Yamada N, Tsumoto K, Ono K, Sutoh H, et al. Redox-coupled alkali-metal ion transport mechanism in binder-free films of Prussian blue nanoparticles. Journal of Materials Chemistry A 2019, 7(9): 4777-4787.DOI
  40. [40]Ashby DS, DeBlock RH, Lai C-H, Choi CS, Dunn BS. Patternable, Solution-Processed Ionogels for Thin-Film Lithium-Ion Electrolytes. Joule 2017, 1(2): 344-358.DOI
  41. [41]Keene ST, Melianas A, van de Burgt Y, Salleo A. Mechanisms for Enhanced State Retention and Stability in Redox-Gated Organic Neuromorphic Devices. Adv Electron Mater 2019, 5(2): 1800686.
  42. [42]Melianas A, Quill TJ, LeCroy G, Tuchman Y, Loo Hv, Keene ST, et al. Temperature-resilient solid-state organic artificial synapses for neuromorphic computing. Sci Adv 2020, 6(27): eabb2958.DOI
  43. [43]Xiao TP, Bennett CH, Feinberg B, Agarwal S, Marinella MJ. Analog architectures for neural network acceleration based on non-volatile memory. Applied Physics Reviews 2020, 7(3): 031301.DOI
  44. [44]Xiang Y, Huang P, Zhao Y, Zhao M, Gao B, Wu H, et al. Impacts of State Instability and Retention Failure of Filamentary Analog RRAM on the Performance of Deep Neural Network. IEEE Transactions on Electron Devices 2019, 66(11): 4517-4522.DOI
  45. [45]Xiao TP, Feinberg B, Bennett CH, Prabhakar V, Saxena P, Agrawal V, et al. On the accuracy of analog neural network inference accelerators. arXiv preprint arXiv:210901262 2021.DOI
  46. [46]Yu S. Neuro-inspired computing with emerging nonvolatile memorys. Proceedings of the IEEE 2018, 106(2): 260-285.DOI
  47. [47]Marinella MJ, Agarwal S, Hsia A, Richter I, Jacobs-Gedrim R, Niroula J, et al. Multiscale Co-Design Analysis of Energy, Latency, Area, and Accuracy of a ReRAM Analog Neural Training Accelerator. IEEE Journal on Emerging and Selected Topics in Circuits and Systems 2018, 8(1): 86-101.DOI
  48. [48]Zhang C, Zhou P. A Quantized Training Framework for Robust and Accurate ReRAM-based Neural Network Accelerators. 2021 26th Asia and South Pacific Design Automation Conference (ASP-DAC); 2021 18-21 Jan. 2021; 2021. p. 43-48.DOI
  49. [49]Agarwal S, Plimpton SJ, Hughart DR, Hsia AH, Richter I, Cox JA, et al. Resistive memory device requirements for a neural algorithm accelerator. 2016 International Joint Conference on Neural Networks (IJCNN); 2016 24-29 July 2016; 2016. p. 929-938.DOI
  50. [50]Duda RO, Hart PE. Pattern classification and scene analysis, vol. 3. Wiley New York, 1973.
  51. [51]Garris MD, Blue JL, Candela GT. NIST form-based handprint recognition system (release 2.0). 1997.DOI
  52. [52]Candela MGIBG, Grother DDIGP, Janet S, Wilson C. NIST form-based handprint recognition system: Technical Report NISTIR 5469, Nat'l Inst. of Standards and Technology; 1994.
  53. [53]Xiao H, Rasul K, Vollgraf R. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:170807747 2017.arxiv.org/abs/1708.07747
  54. [54]LeCun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to document recognition. Proceedings of the IEEE 1998, 86(11): 2278-2324.
  55. [55]Fuller EJ, Li Y, Bennet C, Keene ST, Melianas A, Agarwal S, et al. Redox transistors for neuromorphic computing. IBM Journal of Research and Development 2019, 63(6): 9:1-9:9.DOI
  56. [56]Li Y, Fuller EJ, Sugar JD, Yoo S, Ashby DS, Bennett CH, et al. Filament-Free Bulk Resistive Memory Enables Deterministic Analogue Switching. Adv Mater 2020, 32(45): 2003984.DOI
  57. [57]Agarwal S, Garland D, Niroula J, Jacobs-Gedrim RB, Hsia A, Heukelom MSV, et al. Using Floating-Gate Memory to Train Ideal Accuracy Neural Networks. IEEE Journal on Exploratory Solid-State Computational Devices and Circuits 2019, 5(1): 52-57.arxiv.org/abs/1901.10570
  58. [58]Kuzum D, Yu S, Philip Wong HS. Synaptic electronics: materials, devices and applications. Nanotechnology 2013, 24(38): 382001.DOI
  59. [59]Gualandi I, Tonelli D, Mariani F, Scavetta E, Marzocchi M, Fraboni B. Selective detection of dopamine with an all PEDOT: PSS organic electrochemical transistor. Scientific reports 2016, 6(1): 1-10.DOI
  60. [60]Bignozzi CA, Roffia S, Scandola F. Intervalence transfer in cyano-bridged bi- and trinuclear ruthenium complexes. Journal of the American Chemical Society 1985, 107(6): 1644-1651.DOI
  61. [61]Pieslinger GE, Albores P, Slep LD, Coe BJ, Timpson CJ, Baraldo LM. Communication between Remote Moieties in Linear Ru–Ru–Ru Trimetallic Cyanide-Bridged Complexes. Inorg Chem 2013, 52(6): 2906-2917.DOI
  62. [62]Lin J-L, Tsai C-N, Huang S-Y, Endicott JF, Chen Y-J, Chen H-Y. Nearest- and Next-Nearest-Neighbor Ru(II)/Ru(III) Electronic Coupling in Cyanide-Bridged Tetra-Ruthenium Square Complexes. Inorg Chem 2011, 50(17): 8274-8280.DOI
  63. [63]Robin MB, Day P. Mixed Valence Chemistry-A Survey and Classification. In: Emeléus HJ, Sharpe AG (eds). Advances in Inorganic Chemistry and Radiochemistry, vol. 10. Academic Press, 1968, pp 247-422.
  64. [64]Newton MD. Quantum chemical probes of electron-transfer kinetics: the nature of donor-acceptor interactions. Chem Rev 1991, 91(5): 767-792.DOI
  65. [65]Arnett DC, Voehringer P, Scherer NF. Excitation Dephasing, Product Formation, and Vibrational Coherence in an Intervalence Charge-Transfer Reaction. Journal of the American Chemical Society 1995, 117(49): 12262-12272.DOI
  66. [66]Cammarata M, Zerdane S, Balducci L, Azzolina G, Mazerat S, Exertier C, et al. Charge transfer driven by ultrafast spin transition in a CoFe Prussian blue analogue. Nature Chemistry 2021, 13(1): 10-14.DOI
  67. [67]Barlow K, Johansson JO. Ultrafast photoinduced dynamics in Prussian blue analogues. Phys Chem Chem Phys 2021, 23(14): 8118-8131.DOI
  68. [68]England SJ, Kathirgamanathan P, Rosseinsky DR. Perturbation calculation from the charge-transfer spectrum data of intervalence site-transfer D.C. conductivity in Prussian Blue. J Chem Soc, Chem Commun 1980(17): 840-841.DOI
  69. [69]Austin IG, Mott NF. Polarons in crystalline and non-crystalline materials. Advances in Physics 1969, 18(71): 41-102.
  70. [70]Wei Q, Mukaida M, Naitoh Y, Ishida T. Morphological Change and Mobility Enhancement in PEDOT:PSS by Adding Co-solvents. Adv Mater 2013, 25(20): 2831-2836.DOI
  71. [71]Wojdeł JC, Moreira IdPR, Bromley ST, Illas F. On the prediction of the crystal and electronic structure of mixed-valence materials by periodic density functional calculation: The case of Prussian Blue. The Journal of Chemical Physics 2008, 128(4): 044713.DOI
  72. [72]Spitzer WG, Fan HY. Determination of Optical Constants and Carrier Effective Mass of Semiconductors. Physical Review 1957, 106(5): 882-890.DOI
  73. [73]Hurlbutt K, Giustino F, Pasta M, Volonakis G. Electronic Structure and Electron-Transport Properties of Three Metal Hexacyanoferrates. Chem Mater 2021, 33(17): 7067-7074.DOI
  74. [74]Foster ME, Sohlberg K, Allendorf MD, Talin AA. Unraveling the Semiconducting/Metallic Discrepancy in Ni3(HITP)2. The Journal of Physical Chemistry Letters 2018, 9(3): 481-486.DOI
  75. [75]Furutsuki S, Chung S-C, Nishimura S-i, Kudo Y, Yamashita K, Yamada A. Electrochromism of LixFePO4 Induced by Intervalence Charge Transfer Transition. The Journal of Physical Chemistry C 2012, 116(29): 15259-15264.DOI
  76. [76]Li Y, Fuller EJ, Asapu S, Agarwal S, Kurita T, Yang JJ, et al. Low-Voltage, CMOS-Free Synaptic Memory Based on LiXTiO2 Redox Transistors. ACS Applied Materials & Interfaces 2019, 11(42): 38982-38992.DOI
  77. [77]Fischer ER, Hansen BT, Nair V, Hoyt FH, Dorward DW. Scanning electron microscopy. Curr Protoc Microbiol 2012, Chapter 2: Unit2B.2-2B.2.
  78. [78]Kresse G, Furthmüller J. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Physical review B 1996, 54(16): 11169.pubmed.ncbi.nlm.nih.gov/9984901
  79. [79]Ganose AM, Park J, Faghaninia A, Woods-Robinson R, Persson KA, Jain A. Efficient calculation of carrier scattering rates from first principles. Nature Communications 2021, 12(1): 2222.DOI
  80. [80]Madsen GKH, Carrete J, Verstraete MJ. BoltzTraP2, a program for interpolating band structures and calculating semi-classical transport coefficients. Comput Phys Commun 2018, 231: 140-145.
  81. [81]Hinuma Y, Pizzi G, Kumagai Y, Oba F, Tanaka I. Band structure diagram paths based on crystallography. Computational Materials Science 2017, 128: 140-184.DOI
  82. [82]Momma K, Izumi F. VESTA 3 for three-dimensional visualization of crystal, volumetric and morphology data. J Appl Crystallogr 2011, 44(6): 1272-1276.DOI

Paper details

vlnc:W_75w4pe6h
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents