Organic Synaptic Transistors with an Ultra‐Short‐Term Weight‐Reconstruction for Processing Multiple Types of Signals

An organic synaptic transistor (OST) that utilizes a multiple‐annealing process and an ion gate to achieve significant reductions in operating voltage and increases in transconductance is demonstrated. The OST exhibits an ultra‐short‐term plasticity (USTP) with a maximum retention time of only 20.7 ms, which does not increase with the number and duration of spikes. This is the shortest retention time yet achieved by an ion‐gel‐regulated synaptic transistor. In addition, OST‐integrated array exhibits a tunable weight plasticity and a short weight refresh time for a stable image resetting in ample time of <0.2 s. It is also sensitive to the frequency and amplitude of electrical inputs; a low‐frequency suppression and a nonlinear amplitude gain enables OST‐constructed filter for use in processing multiple types of signals. This work is a step toward constructing high‐performance and multifunctional artificial intelligent systems.


Introduction
Traditional signal processing methods based on von Neumann architecture or Boolean logic rely on physically separated memory and processing modules, therefore cannot efficiently handle DOI: 10.1002/aelm.20230070217] Most reported artificial synapses focused on the diversity of synaptic plasticity and weight reconfigurability, while neglecting the fact that their operating sensitivity may be limited by high operating voltage and low transconductance.Furthermore, these devices cannot maintain rapid weight resetting during the regulation of synaptic weight, especially after receiving reinforcement stimuli; they typically require additional reverse stimuli to inhibit the weight, causing a reduced efficiency and an additional energy consumption.These limitations also result in signal aliasing or loss when processing real-time signals. [16,18]Attempts to employ the interface engineering and dynamically operated electron/ion mechanisms in the synaptic transistor demonstrate the high-efficiency signal processing with rapid reset capabilities, and therefore pave the way to approach advanced intelligent systems for future neuromorphic electronics.
Here, we proposed an ultra-short-term plastic synaptic transistor for processing multiple types of signals.By utilizing an multiple-annealing process and combining an ion gate, the operating voltage significantly decreased while the transconductance increases.The device exhibited a tunable weight plasticity and a fast weight refresh rate, with a maximum retention time of only 20.7 ms, which did not increase even when processing more spikes over longer durations.This represents the best result achieved so far for an artificial synaptic transistor.An integrated array demonstrated a stable image-refresh process with a short reset time of <0.2 s.It also displayed essential short-term behaviors under electrical operation, including paired-pulse facilitation and amplitude-frequency characteristics, which were used for low-frequency suppression filtering and noise reduction.This work provides a new approach to building advanced neuromorphic computing devices.

Results and Discussion
The as-prepared OST device (Figure 1a) consists of an interfaceoptimized organic channel and an ion gel (Figure 1b; Figure S1, Supporting Information).A heavily p-doped silicon wafer with a 300 nm thick thermally oxidized SiO 2 layer served as the dielectric layer/bottom gate for a traditional organic transistor (OT).The silicon wafer was washed sequentially with deionized water, ethanol, acetone, and isopropanol, followed by treatment with ultraviolet ozone for 20 min.A poly(methyl methacrylate) (PMMA) (0.4 wt.%): dioctylbenzothienobenzothiophene (C8-BTBT) (2 wt.%) mixed solution in chlorobenzene was spin-coated onto the Si/SiO 2 layer at 4000 rpm for 30 s.The prepared wafer was solvent-vapor annealed with 120 °C chlorobenzene vapor for 1 h to separate the lower PMMA and upper C8-BTBT films.A second annealing process was conducted in a nitrogen (N 2 ) atmosphere at 60 °C.Source/drain electrodes of 60 nm thick Au were thermally evaporated at a rate of 0.2 Å s −1 through a shadow mask.Finally, the ion gel, composed of the polymer poly(vinylidenefluoride-co-hexafluoropropylene) [PVDF-HFP] and the ionic liquid 1-ethyl-3methylimidazolium bis(trifluoromethylsulfonyl)imide [EMIM-TFSI], was transferred onto the channel area.A W probe that contacted the ion gel was used as the top gate to apply the presynaptic spike.To enable comparison, we fabricated four additional device types, including an OT without solvent-vapor annealing, an OT without the second annealing process, an OST without solvent-vapor annealing, and an OST without the second annealing process.The electrical characteristics of all the OT and OST devices were measured at room temperature using a Keithley 4200CS semiconductor analyzer.
After undergoing two annealing processes, a C8-BTBT upper film that had the typical Stranski-Krastanov growth mode separated from the PMMA: C8-BTBT mixed layer (Figure 1c,d); the thicknesses of the separated PMMA and C8-BTBT films were measured by atomic force microscope (AFM) to be 16 and 54 nm, respectively (Figure S2, Supporting Information).Figure 1e shows X-ray diffraction (XRD) pattern and optical microscope (OM) image of the PMMA: C8-BTBT film: The PMMA: C8-BTBT interface proved to be continuous and uniform, and the addition of PMMA into the C8-BTBT solution directs the formation of organic plate-like crystals by preventing C8-BTBT molecules from strong molecular interactions to form small dots. [19]The XRD signal from the PMMA: C8-BTBT film further confirms that the molecules are packed in the same structure as previously reported. [20]igure 2a shows the typical transfer characteristic curve of a bottom-gate OT at a negative drain-source voltages V DS of −50 V.This OT device exhibits p-type operate-mode behaviors with a weak memory effect.The linear and saturation regions can be clearly observed with the increase of V DS .Despite achieving a drain-source current I DS of −602.2 μA with a negligible weak leakage current, it requires correspondingly high gate-source voltage V GS and threshold voltage V th , with both reaching values as high as −50 and −28 V, respectively (Figure 2b; Figure S3a, Supporting Information).As a result, the calculated transconductance G m is only 15 μS.To improve the ability to regulate the channel with V GS and to store charge, an ion gel was introduced.Figure 2c shows the typical transfer characteristic curve of the OST with top gate driving at a small V DS of only −1 V.The V th decreases sharply to −2.6 V, which is only ≈1/10 of that under the bottom gate driving mode; despite the increase in leakage current caused by the introduction of ion gel, it only accounts for 4.5% of the working current, having no significant impact on the normal operation of the device (Figure S3b, Supporting Information).Thus, the G m was enhanced to 4.9 S based on the output characteristics curve (Figure 2d).That means the electrical double layer formed between the active layer and the ion gel can efficiently induce a large number of charge carriers. [21,22]Due to the continuous accumulation of anions for inducing hole carriers in the channel interface under the influence of a negative electric field, the current remains in an increasing state for a short period of time without reaching saturation.The memory effect in this mode is obvious, which indicates the OST device can process information reliably.
Additionally, parallel experiments provided that the performance of OT and OST devices significantly degraded when either of the two annealing stages was missing, as evidenced by a decrease in transconductance and a increasing driving voltage (Figure S4, Supporting Information).It attributed to an inadequate separation of PMMA/C8-BTBT layer and a rough interface morphology, resulting in a decreased efficiency of hole carrier transport. [23]igure 3a shows the schematic diagram of the OST with top gate driving when regarded as an artificial synapse.The ion gel mimics the synaptic cleft.When no presynaptic spike was applied, anions and cations in the ion gel were distributed randomly (Figure S5a, Supporting Information).When negative presynaptic spikes were applied, anions accumulated at the ion gel/C8-BTBT interface to attract additional holes in the conductive channel to form postsynaptic current (PSC) (Figure S5b, Supporting Information).The source and drain electrodes in the OST function as the post-synaptic membrane to output these postsynaptic responses in real time.Here, a negative spike of −3.5 V was applied to the ion gel for 51 ms, the PSC significantly increased, which sharply decayed to the initial state once the spike was removed (Figure 3b).To quantify the USTP, the PSCretention curves were fitted using a simple first-order stretched exponential function (SEF) [24] : where t 0 = 0 is the time that the presynaptic spike finishes, I 0 is the initial PSC when the spike finishes, I ∞ is the final PSC in an equilibrium state, 0 ≤  ≤ 1 is a stretch index, and  is the reten-tion time.27][28][29][30][31][32][33][34][35] In biology, repeated rehearsal processes can strengthening the synaptic weight. [36,37]Analogously, the PSC under the electric spikes can be lifted by increasing the number N and duration D (Figure 3c): When N (from 1 to 100) and D (from 51 to 512 ms) were increased, the number of anions that migrated increased, and the induced changes in conductivity of OST increased, thereby spike-number-dependent plasticity (SNDP) and spike-duration-dependent plasticity (SDDP) and increased proportionally (Figure 3d,e).It is noteworthy that  consistently remains an ultra-low level of below 20 ms with the increase of synaptic weights during the processes, indicating that the OST device has a strictly USTP (Figures S6 and S7, Supporting Information).This phenomenon is due to the abundance of shallow traps in active layer, which shortens the transmission life of hole carriers, which ensures both the instantaneous capture and the rapid release of the charges (Figure S5c, Supporting Information). [38]Furthermore, since the applied spike voltage is only slightly higher than the V th of the device, the built-in potential for keeping the hole carrier transport collapses once captured charges are released.As a result, the PSC degradation is particularly significant.
The OST device was subjected to a series of cyclic tests in nitrogen and air environments to analyze the electrical stability (Figure S8, Supporting Information).The output current in the first 100 pulsed cycles is almost unchanged.After exceeding 1000 cycles, a slight decline in peak current was observed, but even at 6000 cycles, the PSC could still maintain over 89% of its initial value.Subsequently, similar tests were conducted in an air environment, revealing an accelerated rate of current decay.However, even after 4000 cycles of spikes, the current remained above 80% of its initial value.This relative tolerance for cyclic testing and variable environments demonstrates that OST has potential application in electronics.
Associating USTP with weight-mapped pictures mimics the image-refresh process in consequence of the volatile behaviors of array devices.During events #1-8, the image transition from the letter "O" to the letter "S" and then to the letter "T" was displayed on 5 × 5 synaptic units by mimicking a letter image composed of 25 pixels (Figure 4a-h).By using a simple perceptron, the three types of letters were classified, and the weight ranges for each classification label were obtained (Figure 4i): a single lowwidth (i.e., D = 0.05 s) pulse is sufficient to regulate the weights to reach the range of target letter, which can be quickly reset within 0.05 s without affecting the input of the next set of signals; under stronger pulses of large-quantity (i.e., N = 100) or high-width (i.e., D = 0.5 s), the reset speed may be affected, and the weight can still recover to the initial value within 0.2 s.
Paired-pulse facilitation (PPF) is a typical short-term facilitation phenomenon, in which the latter PSC is larger than the former when two successive pulses arrive. [39,40]Analogous to biological neural, PPF function was emulated using our OST device: a pair of spikes of −3.5 V with different intervals Δt evokes a paired current response (Figure 5a).We calculated PPF index by using the gain as induced by the second spike (A 2 )/that induced the first spike (A 1 ).This decay curve fitted well with a biexponential function, and was comparable to the data in biological synapses [41] : The initial facilitation magnitudes are given by C 1 and C 2 .The two timescales  1 and  2 are 10.3 and 359.7 ms, representing characteristic relaxation time of rapid and slow phase.The value of the  1 approximately aligns with the fitted  of PSC decay, providing further evidence for the rapid decay of the PPF index within the ultra-short-term time range.Once Δt is larger than  2 , the PSC increment evoked by the second spike is limited, resulting in the PPF index gradually approaching to 100% with the increase in Δt.The values of both  1 and  2 limited to the millisecond scale, are consistent with the ultra-short-term characteristics of the OST device.
The synaptic weight can be changed intuitively by regulating the frequency of input spikes, which is defined as spikefrequency-dependent plasticity (SFDP). [42]The PSCs in response to the successive spikes with different frequencies were recorded (Figure 5b): With the frequency decreasing, PSC also decreases gradually, due to reduced ion migration and the increased possibility of ions back-diffusing in the OST when the spikes are removed.When the spike frequency reduces to a lower value, e.g., 0.32 Hz, the balance between ions migration and backdiffusion is achieved, resulting in relatively stable PSCs with a negligible current gain (Figure S9, Supporting Information).The dependence of current gain on peak frequency was well fitted with the sigmoidal function, which was analogous to the lowfrequency suppression filter characteristic observed in biology (Figure 5c). [43]That indicates the input signals with high frequency exceeding a certain cut-off value f c are allowed to pass through OST devices, while low frequency signals are immensely weakened.Figure 5d depicts the schematic illustration of OST devices as low-frequency suppression filters by creating an image of flowers as an example for simulating the filtering process: when the f c = 2.32 Hz was utilized for the low-frequency suppression filtering, the image outline characteristic could be sharpened.
Moreover, an amplitude filter can also be achieved by applying spikes with different amplitudes.With the spike voltage increasing from −0.5 to −3.5 V, OST exhibited a nonlinear spike-voltagedependent plasticity (SVDP) index (Figure 5e). [44]The inset in Figure 5e depicts the input-output conversion relationship of OST-based filters for signal processing.In order to show the OST-based filtering function, three images with different noisy background were treated to simulate the filtering process (Figure 5f).When the grayscale mapping between input and ouput was utilized for filtering, the images could be denoised, which revealed a marked improvement compared to the original images.

Conclusion
In summary, we presented a transistor-structured synaptic device capable of ultra-short-term weight-reconstruction for processing multiple types of signals.The OST device exhibited excellent performance in emulating biological synapses, with a retention time of only 20.7 ms, which did not increase with the number and duration of spikes.This is the shortest retention time yet achieved by an ion-gel-regulated synaptic transistor.The device also mimiced a stable image-refresh process with a short reset time of <0.2 s.Additionally, the device demonstrated the ability to function as a low-frequency suppression filter and a nonlinear amplitude filter for signal processing.The results provide a promising direction for the development of neuromorphic computing and artificial intelligence systems.

Figure 1 .
Figure 1.a) Schematic diagrams of biological neuron and OST device.b) Fabrication of the C8-BTBT-based OT and OST devices.c) AFM image of the PMMA film.d) AFM image of the PMMA: C8-BTBT film.e) XRD pattern of the PMMA: C8-BTBT film, and OM photograph of the PMMA: C8-BTBT film (inset).

Figure 2 .
Figure 2. a) Transfer characteristic curve of the C8-BTBT-based OT. b) Output current-voltage characteristics of the C8-BTBT-based OT. c) Transfer characteristic curve of the C8-BTBT-based OST.d) Output current-voltage characteristics of the C8-BTBT-based OST.

Figure 3 .
Figure 3. a) Schematic diagram of the C8-BTBT-based OST with top gate driving when regarded as an artificial synapse.b) PSC response of the OST measured at a spike of −3.5 V, and PSC-retention curve of OST after the spike (inset).c) PSC changes under spikes of different numbers (A: −3.5 V, N: 5-100, D: 51 ms, interval time: 51 ms) and durations (A: −3.5 V, D: 51-512 ms).d) SNDP index plotted as a function of presynaptic spike number (top), and fitted  values of OST after spikes of different numbers (bottom).e) SDDP index plotted as a function of presynaptic spike duration (top), and fitted  values of OST after spikes of different durations (bottom).

Figure 5 .
Figure 5. a) PPF index is plotted at different interval times between two spikes (A: −3.5 V, D: 51 ms).b) PSC changes under spikes of different frequencies (A: −3.5 V, N: 20, D: 51 ms, f: 0.32-10.45Hz).c) SFDP index plotted as a function of presynaptic spike frequency.d) The image of flowers sharpened with the OST-based filtering function at the cut-off frequency of 2.32 Hz. e) SVDP index under spikes of different voltages (A: 0.5−3.5 V, N: 20, D: 51 ms, f: 1.28 Hz).f) Gray values of input and output images after denoising with OST-based filtering function.