Introduction

Neuromorphic systems that interact with the physical world1,2 emulate neural computation through distributed, event-driven, and low-power electronic systems comprising mixed-signal artificial neuron and synaptic circuits. Each neuron circuit integrates input spikes via its tunable synapses and generates output spikes when its threshold is exceeded, supporting complex behaviors like temporal coding and learning3,4. These systems represent a promising solution for applications involving processing of noisy sensory signals, for embodied robotic intelligence, brain–machine interfaces, environmental monitoring, or wearable bio-signal analysis5,6. To date, with a handful of exceptions, the majority of neuromorphic hardware relies on silicon-based complementary metal-oxide semiconductor (CMOS) circuits7,8,9,10, while their organic counterpart is less mature. Organic electronics offer a compelling route towards biocompatible, flexible, conformable, and stretchable neuromorphic hardware, making them better suited for interaction with soft and fragile materials and biological tissue. Additionally they have the advantages of low-temperature fabrication on large, soft substrates using scalable, low-cost and large-area techniques11,12. Among these, organic electrochemical transistors (OECTs)—which rely on ionic exchange between the electrolyte and the active layer—have been widely adopted for emulating synaptic behavior due to, among other things, their high transconductance, ionic transport, and analog memory properties13,14,15,16,17,18,19.

Despite these advantages, OECTs face critical limitations that constrain their utility in neuromorphic systems, specifically in embodied AI applications, such as integration with robotics. First, their operation depends on ionic transport, which imposes slow response times (typically limited to tens of hertz or less) and makes them unsuitable for fast, dynamic processes such as spike generation or coincidence detection. Second, ionic transport typically relies on aqueous environment, not suitable for many applications. Third, OECTs require continuous ion injection from an electrolyte, which reduces long-term stability and complicates device integration. Next, their active material—often PEDOT:PSS—exhibits cytotoxicity, acidic behavior, and excess negative charge, all of which hinder effective interaction with biological molecules and cells. Furthermore, PEDOT:PSS-based devices operate in depletion mode, requiring relatively high gate voltages and constant doping, which can lead to parasitic reactions with water and oxygen—ultimately degrading performance20,21,22,23.

While organic synaptic devices using ionic and electrochemical mechanisms have been demonstrated24,25,26, most implementations remain limited to passive synaptic emulation and somatic circuits alone27,28,29,30. To the best knowledge of authors only two prior reports have achieved co-integration of synaptic and somatic elements18,24, yet these lack features such as multi-synapse integration or mechanical robustness. Our work addresses this unmet challenge by demonstrating a fully organic, non-ionic spiking neuron that integrates excitatory and inhibitory synapses with a soma on a flexible substrate, enabling biologically relevant signal processing, adaptation, and resilience to mechanical stress. A comparative summary of recent organic neuron implementations is provided in Supplementary Table 1.

To overcome these limitations, we have pursued a fundamentally different approach using organic field-effect transistors (OFETs). OFETs operate via electronic charge transport, enabling faster response, better control, and greater integration potential for spiking neural circuits. In our earlier work, we demonstrated OFET-based non-spiking neuromorphic primitives, including analog synaptic circuits and somas31,32,33. These advances culminated in the development of hybrid NUCLEOs system, a laminated organic–inorganic neuromorphic robot capable of legged locomotion34,35. Recognizing the computational and power consumption advantages of spiking neural networks (SNNs), we then introduced OFET-based circuits for spike-based somas and tunable synapses, including integrate-and-fire Axon-Hillock soma36, Log-Domain Integrator (LDI) and Differential Pair Integrator (DPI) synaptic architectures37,38,39.

In this work, we present the first demonstration of a full organic spiking neuron, fit with tunable excitatory and inhibitory synapses. The neuromorphic circuit is composed of p- and n-channel complementary OFETs and passive capacitive elements all integrated on the same flexible substrate. Our neuromorphic circuit combines the biocompatible properties of organic electronics with the computational power of spiking neurons. It offers tunable synaptic weighting, biologically plausible excitatory and inhibitory time constants, and linear somatic dynamics including frequency modulation, signal summation, and coincidence detection. We further validate its real-time responsiveness in a closed-loop light-control task, demonstrating the feasibility of applying organic spiking neurons in sensory feedback systems. This work establishes a new direction for organic neuromorphic electronics—one that is faster, more stable, and more compatible with biological systems—by demonstrating biologically inspired spiking behavior on a flexible, low-power platform. While the current operating voltages are optimized for circuit-level proof-of-concept, the underlying architecture supports future voltage scaling for safe and practical bio-interfacing, paving the way for embodied AI applications in soft robotics, prosthetics, and implantable neurointerfaces.

Results

Fabrication

The fabrication process for organic field effect transistors (OFETs) and circuit components in this study is based on previously established protocols for organic synaptic circuits37,38,39. The devices were fabricated on flexible glass substrates supplied by Corning Incorporated, which enabled mechanical flexibility while maintaining substrate stability. The process began with thermal evaporation of a 5 nm chromium (Cr) adhesion layer, followed by a 30 nm silver (Ag) electrode layer, and a 20 nm film of aluminum oxide (Al2O3), which serves as a gate dielectric.

To enhance the dielectric interface, the surface of Al2O3 was functionalized using Silane A-175, followed by the deposition of a 100 nm layer of Parylene diX-SR through chemical vapor deposition. Subsequently, a 5 nm self-assembled monolayer (SAM) of Triptycene was applied to improve charge carrier injection and interfacial stability.

The active semiconducting layers were thermally deposited atop the parylene layer. For the p-type OFETs (pOFETs), we employed S-DNTT-10, while the n-type OFETs (nOFETs) used TU-1 (more details can be found in Materials and Methods section). After the semiconductor deposition, multiple 30 nm gold (Au) layers were sequentially evaporated to form the source and drain contacts, the top electrodes of the capacitors, and the interconnects for the circuit architecture.

Layer patterning was achieved through a shadow-mask technique, ensuring reproducibility across multiple device iterations. Via holes were created by the laser ablation technique to expose interlayer contacts and facilitate vertical integration. A cross-sectional illustration of the device stack and a representative photograph of the fabricated OFETs are shown in Fig. 1a, b.

Fig. 1: Characterization of individual OFETs and integrated capacitors.
Fig. 1: Characterization of individual OFETs and integrated capacitors.The alternative text for this image may have been generated using AI.
Full size image

a, b A schematic diagram and a photograph of p-type and n-type OFETs fabricated on a flexible substrate (Scale bar: 500 μm). c, d Output characteristics of p-type and n-type OFETs under flat (solid blue) and tension-bent (dashed red) conditions. Bending enhances mobility and on-current in p-type devices while reducing both in n-type devices. e Capacitance characterization of eight parasitic circuit-integrated capacitors, measured from 1 Hz to 5 kHz. Inset: variance of capacitance across devices as a function of frequency, showing convergence at higher frequencies and indicating good device-to-device uniformity.

OFET and capacitor characterization

Mechanical flexibility is a key advantage of organic electronics. Therefore, all devices and circuits in this study were fabricated on flexible substrates and characterized under both flat and tension-bent configurations (Supplementary Fig. 1). OFETs were designed with a channel length of 50 μm and width of 500 μm, with a 10 μm gate–source/drain overlap. These dimensions were selected based on shadow mask limitations and to minimize parasitic capacitance associated with the gate-source and gate-drain electrodes overlap. Through systematic experimentation, it was found that both p- and n-type OFETs exhibit nearly identical characterizations, justifying the use of the same channel geometry for both devices.

Figure 1c, d show representative characteristic results for p- and n-type OFETs, with flat and tension-bent configurations represented by solid blue and dashed red lines, respectively. As summarized in Table 1, the threshold voltage remained nearly constant for both device types under mechanical strain. However, bending did induce changes in mobility: p-type OFETs exhibited a threefold increase in mobility and on-current, while n-type devices showed a twofold reduction. These changes are attributed to strain-induced morphological changes in the organic semiconductors. In p-type devices, strain likely enhances π–π stacking and chain alignment, improving charge transport40,41. In contrast, n-type materials are more sensitive to molecular disorder and trap formation, resulting in degraded performance under strain42,43.

Table 1 Comparison of threshold voltages and carrier mobilities of p- and n-type, flat and tension-bent OFETs represented in Fig. 1c, d alongside the Standard Deviation (STD) of measurement showing in Supplementary Figs. 4 and 8 for flat configuration
Full size table

Device reliability and variability were assessed through repeated stress testing and statistical sampling. Supplementary Figs. 2 and 3 show that repeated electrical measurements over seven cycles did not significantly affect the transfer characteristics of either p- or n-type OFETs. Device-to-device variability across eight devices on a single chip is shown using whisker plots (Supplementary Figs. 7-9 and 11-13). For both transistor types, the on/off current ratio varied across four to five orders of magnitude, threshold voltage varied within ~150 mV (p-type) and ~120 mV (n-type), and mobility spanned a window of 0.06 cm2/V s. These results confirm a high degree of reproducibility across devices.

All circuit layers were patterned using shadow masking. This approach, while low-cost and compatible with printing processes, imposes resolution constraints that affect circuit design. While the current fabrication uses thermal and vacuum deposition for proof-of-concept, the materials and architecture present a path towards future translation to solution-processable and printed fabrication using appropriate material substitutions. It should be noted, that a notable contribution of this work is the co-integration of thin-film capacitors and OFETs on the same flexible substrate—a rarely demonstrated achievement in organic neuromorphic circuits28. Although we have previously implemented integrated nanofarad capacitors in analog OFET-based circuits37,38, such large capacitances severely limit spiking circuit frequency. Attempts to reduce capacitor area proved insufficient due to the dominance of parasitic gate-source/drain overlap capacitances, which inherently constrain the system’s speed.

To overcome this, we redesigned the architecture to exploit parasitic overlap capacitances as functional circuit elements. The resulting capacitors—fabricated across eight positions on the chip—were characterized using an LCR meter from 1 Hz to 5 kHz (Fig. 1e). At 1 Hz, measured capacitances ranged from 10 to 30 pF, dropping to approximately 1 pF at 5 kHz. This frequency-dependent reduction stems from how LCR meters model capacitance using ideal RC equivalents. At low frequencies, interfacial polarization and slow dielectric responses contribute significantly, inflating the measured capacitance. At higher frequencies, these contributions diminish, revealing the intrinsic high-frequency capacitance of the dielectric stack.

The inset in Fig. 1e displays the frequency dependence of capacitance variance across eight devices. At 1 Hz, the variance reaches up to 25 pF2, reflecting interfacial and leakage-related variability. As frequency increases, variance drops toward zero, suggesting that the observed dispersion is not intrinsic but driven by extrinsic frequency-dependent effects. This convergence supports the conclusion that dielectric properties across devices are fundamentally uniform—a critical quality for reliable neuromorphic circuit integration44.

Organic integrate-and-fire neuron and synaptic circuit characterization

The integrate-and-fire (IF) is a simplified yet widely used neuron model that captures essential neuronal behavior by integrating synaptic inputs over time and generating a spike when a threshold is reached, followed by a reset of the membrane potential. Although complex biological details such as dendritic processing or ion channel dynamics are omitted, its simplicity and computational efficiency make it ideal for large-scale emulations. Among hardware implementations, the Axon-Hillock (AH) circuit represents an early artificial IF neuron with minimal components, yet it has demonstrated functional viability in neural networks36,45. A common hardware approach for mimicking synaptic behavior in such models is the use of pulsed current-source (PCS) synaptic circuits, which deliver brief current pulses in response to pre-synaptic spikes. These circuits emulate the transient nature of biological synapses and, thanks to the slower temporal dynamics of the organic transistors, serve as input integrators for the IF neuron, with the pulse amplitude and duration encoding synaptic weight and temporal dynamics. In traditional silicon implementations, instead, PCS synapses act as weighted instantaneous switches, and the conduction speed makes it impossible to implement the dynamics of the post-synaptic current without resorting to more complex circuits, such as DPI46. This method provides a scalable and energy-efficient solution for organic-based neuromorphic systems, aligning well with the IF model’s assumptions, facilitating the emulation of spatiotemporal processing in artificial neural networks12,13,33,47,48,49,50,51,52,53,54,55.

In this study, the organic spiking neuron comprises a single AH-based somatic circuit coupled with two excitatory and two inhibitory PCS synaptic circuits. A schematic of the complete circuit, along with a corresponding photograph of a fabricated circuit, is shown in Fig. 2a, b. Both images share a consistent color mapping for clarity indicating individual circuit primitives. Fig. 2c displays the fully fabricated chip under mechanical bending, featuring 20 complete neurons and 20 standalone OFETs. It is important to note that the minimum bending radius of 205 mm used in this study represents the mechanical limit of the flexible glass substrate, beyond which substrate fracture occurs. While smaller radii could not be tested due to substrate fracture, device performance remained stable under this constraint, demonstrating robustness to moderate mechanical stress. As described earlier (see OFET and Capacitor Characterization subsection), conventional capacitors were omitted from the circuit and replaced by parasitic overlap capacitances, simplifying the architecture while maintaining functionality.

Fig. 2: Architecture and characterization of the organic spiking neuron.
Fig. 2: Architecture and characterization of the organic spiking neuron.The alternative text for this image may have been generated using AI.
Full size image

a Schematic of the spiking neuron circuit comprising a single AH somatic circuit and two excitatory and two inhibitory PCS synaptic circuits. b Micrograph of the fabricated neuron with color-coded layout matching the schematic (Scale bar: 2000 μm). The overall footprint of the device is 14 mm × 5 mm, sized intentionally to accommodate misalignments inherent to the shadow-mask-based fabrication process used in this work. c Photograph of the full chip under mechanical bending, containing 20 neurons and 20 individual OFETs. d (Top) Frequency-current (F-I) response of the AH circuit, showing the relationship between input current and output firing rate. (Bottom) Power consumption as a function of frequency under flat and tension-bent conditions. e The input membrane and output voltage of the AH circuit in flat and tension-bent conditions.

To evaluate the AH circuit independently of synaptic input, a semiconductor parameter analyzer was used to inject a programmable input current. The spiking behavior of the circuit was quantified by plotting the output firing frequency as a function of the input current, resulting in a frequency–current (F–I) response, shown in Fig. 2d. The circuit initiates output spiking at 10 Hz with a sub-nanoamp input. Increasing the input to 80 nA raises the firing rate to 40 Hz—a 75% relative increase. Beyond 100 nA, the output frequency saturates, reaching a maximum of 55 Hz, resulting in a canonical sigmoidal activation function of the input-output relationship56. Power consumption was computed as the average product of supply current and applied voltage over time. The AH circuit consumed less than 10 nW at baseline, with power consumption increasing with increasing spike frequency. tension-bent circuits consumed more power than flat configurations, likely due to bending-induced threshold shifts in the p-type OFETs. Energy consumption by individual synaptic circuits was approximately an order of magnitude lower than that of the AH soma, as the synaptic OFETs operate in the subthreshold regime. Therefore, total neuron-level power consumption is dominated by the somatic circuit. Fig. 2e shows the output and membrane signals of the AH circuit in flat and tension-bent conditions. Additional data demonstrating continuous spiking of the AH circuit under constant input current for 6 s is provided in Supplementary Fig. 12.

The PCS synaptic architecture is schematically illustrated in Fig. 2a (green and blue boxes). Each excitatory synapse uses two p-type OFETs, while inhibitory ones use two n-type OFETs, corresponding in synaptic current flowing to or away from the somatic input membrane. All devices are ideally biased in the subthreshold regime, where small gate voltages produce exponentially controlled current pulses, ideal for mimicking analog synaptic integration. This operating mode enables low power consumption and analog weight modulation, aligning closely with the physiological dynamics of biological synapses. In practice, due to variability in OFET behavior, some characterization was performed with devices partially biased above the subthreshold regime.

The gate bias applied to the weighing transistor (Vw) modulates the synaptic efficacy. This parameter controls the magnitude and time constant of the postsynaptic current, directly influencing the temporal filtering properties of the synapse. Biological neurons rely on membrane time constants to filter spike trains, and hence analog hardware must reflect this property. Supplementary Figs. 13 and 14 show the extracted time constants of excitatory and inhibitory synapses across various Vw values under both flat and tension-bent conditions. The excitatory PCS circuit achieved a maximum time constant of 65 ms at Vw = 6.5 V, while the inhibitory circuit reached 275–280 ms at Vw = 0.4 V. These values remain relatively unaffected by mechanical bending. We have detailed the measurement method in prior publications37,38

Figure 3a, b illustrate how Vw modulates synaptic signal amplitude. In excitatory synapses (Fig. 3a), increasing Vw reduces the amplitude of the output, while in inhibitory synapses higher Vw increases the absolute value of the output. Marginal overshoot artifacts appear when the current direction through the OFET reverses, producing hysteresis-related spikes. Interestingly, bending exhibits opposing effects on the two types of synapses: it enhances signal amplitude in excitatory circuits but attenuates it in inhibitory ones. This divergence is attributed to the distinct strain responses of p- and n-type OFETs—bending enhances mobility in p-type semiconductors while degrading it in n-type, consistent with single-device behavior under strain.

Fig. 3: Synaptic signal modulation by weighing voltage in excitatory and inhibitory PCS circuits.
Fig. 3: Synaptic signal modulation by weighing voltage in excitatory and inhibitory PCS circuits.The alternative text for this image may have been generated using AI.
Full size image

a Response of the excitatory synaptic circuit under flat and tension-bent conditions. As the weighing voltage (Vw) increases, the amplitude of the excitatory postsynaptic signal attenuates. b Response of the inhibitory synaptic circuit under flat and tension-bent configurations. In contrast to the excitatory case, increasing Vw enhances the absolute amplitude of the inhibitory signal.

The spiking neuron’s frequency response

A spiking neuron is generally characterized by its firing rate in response to a range of pre-synaptic pulse train frequencies57,58,59. Figure 4a, b present fitted curves to the measured data, showing the relationship between pre-synaptic input and post-somatic output firing rates under eight different synaptic weight voltages, for both excitatory and inhibitory synaptic circuits, in flat and tension-bent conditions—depicted by solid and dotted lines, respectively. The pre-synaptic frequency ranges from 1 to 100 Hz (with pulses of 5 ms duration), resulting in output frequencies that rise from 0 to 60 Hz for excitatory synapses and attenuate from 60 to 0 Hz for inhibitory ones, depending on the strength of the synaptic weight Vw. The raw data points corresponding to each condition are provided in Supplementary Figs. 15 to 18. The resulting frequency response exhibits a sigmoidal profile (bounded output for unbounded input), highlighting the nonlinear dynamics of the integrate-and-fire neuron: at low input frequencies, insufficient synaptic integration limits output spiking, while a sharp increase in firing occurs once the threshold is surpassed, followed by saturation due to recovery time or refractory constraints. The synaptic weight voltage effectively modulates this response; in excitatory circuits, decreased weighting enhances firing rates, while in inhibitory circuits, higher weights suppress spiking, eventually silencing the neuron at maximum bias. This tunability allows for fine-grained rate control, analogous to biological excitatory and inhibitory modulation. Notably, mechanical bending improves the function of both excitatory and inhibitory synaptic circuits, likely due to enhanced charge transport and improved coupling in the organic semiconductor layers, which results in more efficient current delivery and sharper neural modulation. These results underscore the system’s capability for biologically inspired temporal coding and adaptive signal processing in neuromorphic applications.

Fig. 4: Characterization of the spiking neuron’s frequency response, linearity, and coincidence detection.
Fig. 4: Characterization of the spiking neuron’s frequency response, linearity, and coincidence detection.The alternative text for this image may have been generated using AI.
Full size image

Firing rate of the organic spiking neuron as a function of input frequency for eight synaptic weight voltages under flat (solid) and tension-bent (dotted) conditions, for a excitatory and b inhibitory PCS synaptic circuits. The sigmoidal response demonstrates input-output nonlinearity characteristic of integrate-and-fire dynamics, modulated by synaptic weight and mechanical bending. c Linearity assessment using two excitatory synaptic inputs with fixed frequency and weight. Top: the train of pre-synaptic signal at 10 Hz. upper-middle and middle: neuron response to individual activation of synapse 1 and synapse 2. Lower-middle: response to simultaneous activation of both synapses. Bottom: Whisker plots indicate that the combined firing frequency (1.26 Hz) closely matches the linear sum of individual inputs (0.6 and 0.66 Hz), validating the neuron’s ability to linearly integrate temporally concurrent signals. d Coincidence detection characterization with two subthreshold presynaptic signals applied with increasing inter-stimulus intervals (ISIs). The red trace shows membrane potential dynamics during integration. Full dataset is available in Supplementary Fig. 19.

Linearity of signal integration

Linearity plays a crucial role in simplifying computational processes by establishing a direct, proportional relationship between neuron’s output and the weighted sum of its inputs. This behavior not only streamlines signal processing but also enhances the interpretability and predictability of the network’s response. Linear operations are mathematically tractable and compatible with well-established analytical methods such as regression, convolution, and transformation, which are foundational to fields including machine learning, control systems, and data analysis60,61,62,63,64. Consequently, the presence of linearity within the spiking neurons contributes to improved computational efficiency, network transparency, and scalability.

Figure 4c demonstrates the linear integration capability of the organic spiking neuron. In this experiment, which follows an earlier device configuration detailed in the author’s dissertation39, two excitatory synaptic circuits were connected to a single AH somatic circuit. Both the weighing voltage and the presynaptic signal frequency were held constant as shown in the top inset of Fig. 4c. First, a presynaptic spike train was applied independently to synaptic circuit 1, and the resulting output frequency of the neuron was recorded, as shown in the upper-middle of Fig. 4c. The same test was repeated with synaptic circuit 2 under identical conditions, and its result is shown in the middle.

To assess linearity, both presynaptic signals were simultaneously applied to the two synaptic circuits (lower-middle of Fig. 4c). The neuron’s output frequency was again measured and compared to the sum of the individual responses. Whisker plots at the bottom of Fig. 4c summarize the measured firing rates. Synaptic circuits 1 and 2 yielded mean output frequencies of 0.6 Hz and 0.66 Hz, respectively. When both signals were simultaneously applied with matched phases (i.e., zero phase difference), the neuron’s firing frequency ranged between 1.25 Hz and 1.27 Hz (25th to 75th percentile), demonstrating close agreement with the linear sum of individual inputs. It is important to note that the observed linearity is contingent on both the input frequency and their relative phase. In this study, both presynaptic spike trains were synchronized to ensure temporal alignment, thereby maximizing summation and minimizing destructive interference. Variations in relative timing (phase difference) would likely affect the summation behavior and could lead to sublinear or supralinear integration, depending on the degree of temporal overlap.

While minor deviations were observed-attributable to device-to-device variability-these results confirm the neuron’s ability to linearly sum synaptic inputs. This linear behavior supports reliable spatiotemporal processing in network configurations, reinforcing the viability of OFET-based organic neurons for spiking neuromorphic computing systems.

Coincidence detection

Coincidence detection is a key spiking neural computation, a key differentiator from non-spiking networks, that enables a neuron to recognize and respond selectively when multiple synaptic inputs arrive within a narrow time window. This mechanism mirrors temporal summation in biological neurons, where the precise alignment of presynaptic spikes determines whether an action potential is generated65,66,67,68,69.

In artificial spiking systems, coincidence detection enhances the network’s sensitivity to spatiotemporal input patterns. This temporal gating capability allows neurons to distinguish synchronous events from asynchronous noise, improving performance in tasks such as pattern classification, event-based feature extraction, and motion perception57,70,71,72,73. By integrating temporally aligned spikes across multiple inputs, artificial neurons can implement biologically inspired computations with high efficiency and specificity.

Figure 4d illustrates a summary of the experimental demonstration of coincidence detection in the organic spiking neuron, consistent with the previous linearity experiment configuration. Two presynaptic square-wave inputs (25 ms duration), each producing insufficient input current to trigger an output spike, were delivered with incrementally increasing inter-stimulus intervals (ISIs) from start to start, each defined as a multiple of 34.7 ms. When the ISI was short (e.g., <243 ms), the synaptic inputs effectively summed within the membrane to trigger an output spike. At an ISI of 242.9 ms, the neuron’s response was noticeably diminished, and by 260.25 ms, the neuron failed to spike, indicating a breakdown in temporal summation.

The membrane voltage trace (red line) provides a clear visualization of the integrative dynamics in the somatic circuit, highlighting how synchronized inputs result in suprathreshold activity, while desynchronized inputs fall below the spiking threshold. A comprehensive dataset for this experiment is shown in Supplementary Fig. 19, further validating the neuron’s role in temporally precise signal integration.

Short-term adaptation and synaptic modulation

Adaptation is a critical feature of artificial neural systems, enabling real-time adjustment to changing environmental conditions. While traditional learning in artificial networks involves long-term updates to synaptic weights based on accumulated data and task performance, shorter timescale modulation—often referred to as short-term plasticity or adaptation—can allow systems to dynamically respond to transient stimuli without persistent memory encoding62,63,64,74,75.

In neuromorphic hardware, such short-term adaptation is often achieved through the temporary adjustment of synaptic weights in response to immediate environmental feedback. While mechanisms like long-term potentiation (LTP) and depression (LTD) underlie enduring changes in synaptic strength, transient modulation mechanisms serve as a computationally efficient alternative for real-time control and stimulus tracking57,76,77,78.

Figure 5a, b illustrate an experimental demonstration of short-term adaptive modulation in an organic spiking neuron embedded in a closed-loop light regulation system. The neuron was integrated into a feedback loop comprising a VEML 7700 ambient light sensor and an Arduino Mega 2560 microcontroller. The system regulated LED brightness based on ambient light levels by adjusting the neuron’s synaptic weight and output spike frequency. Under dim lighting, the neuron reduced its spike output and synaptic strength, while bright conditions triggered an increase in both parameters. Fig. 5a depicts the full setup, and Fig. 5b shows membrane voltage, spiking activity, and microcontroller data in low and high light conditions. Although the system modifies synaptic weight in response to sensory input, this modulation is temporary and resets once the environmental condition stabilizes—suggesting a form of short-term adaptation rather than persistent learning. Supplementary videos demonstrate the system’s responsiveness to dynamic lighting. These findings show how OFET-based spiking neurons can support adaptive behavior in neuromorphic control systems, even in the absence of long-term memory storage.

Fig. 5: Demonstration of feedback control loop in an organic spiking neuron via closed-loop light control.
Fig. 5: Demonstration of feedback control loop in an organic spiking neuron via closed-loop light control.The alternative text for this image may have been generated using AI.
Full size image

a Experimental setup showing the organic spiking neuron integrated with a VEML 7700 ambient light sensor and an Arduino Mega 2560 processor in a real-time feedback loop. The neuron modulates its output spike frequency based on ambient light intensity to control the luminance of an LED. b Dynamic response of the neuron under dark and bright environmental conditions. From top to bottom: recorded membrane voltage, generated action potentials. Under darker conditions, synaptic weight and spike frequency decrease, dimming the LED. Conversely, brighter conditions induce an increase in both, brightening the LED output.

Although the modulation observed here does not result in persistent long-term memory, it closely mimics short-term plasticity observed in biological systems, such as short-term depression and facilitation. These transient adaptive mechanisms are essential in sensory gating, dynamic filtering, and real-time attention control in the brain79,80. By temporarily adjusting synaptic weights and spiking activity in response to environmental stimuli, our circuit emulates similar biologically inspired dynamics.

Discussion

In this work we present the first fully organic spiking neuron fit with multiple excitatory and inhibitory synapses, fabricated entirely from OFET-based circuitry with integrated capacitors on the same flexible substrate. By leveraging parasitic capacitors, subthreshold pulsed current-source synaptic circuits, and an integrate-and-fire Axon-Hillock-inspired soma, the neuron demonstrates biologically relevant behaviors, including frequency modulation, signal integration, coincidence detection, and adaptive learning. Characterization under both flat and tension-bent conditions confirms mechanical robustness. Integration with sensory feedback showcases embodied AI application with real-time learning in an ambient light control task.

Importantly, all devices were fabricated using shadow masking techniques without relying on photolithography—essentially outside of a clean room environment—yet still achieved spiking activity at biologically plausible frequencies. This highlights the accessibility and scalability of the platform. With the future implementation of higher-resolution photolithographic patterning, the system’s performance could be further enhanced into the kilohertz regime, comparable to native spiking rates of biological neurons. These findings establish a promising foundation for energy-efficient, biocompatible neuromorphic systems and pave the way for flexible, implantable, and wearable brain-inspired electronics.

Methods

Materials and equipment

The p-type organic semiconductor, 3,10-Didecylnaphtho[2,1-b]naphtho[1’,2’:4,5]thieno[2,3-d]thiophene (S-DNTT-10), and the n-type semiconductor, 4,8-Bis[5-[3-(trifluoromethoxy)phenyl]-2-thienyl]-2λ4δ2-benzo[1,2-c:4,5-c’]bis[1,2,5]thiadiazole (TU-1), were procured from Tokyo Chemical Industry Co., Ltd. Parylene diX-SR was obtained from Daisan Kasei and deposited using a chemical vapor deposition (CVD) process via the SCS LabCoater 3 (PDS 2010). Al2O3 was deposited using Cambridge Nanotech Fiji ALD.

Metal source materials—including chromium (Cr) rods, gold (Au), and silver (Ag) pellets—were supplied by Kurt J. Lesker Company (KJLC). Thin film deposition of gate, source, drain, and electrode layers was performed using a NANO 36 thermal evaporator (KJLC). Electrical characterization, including current-voltage (I–V) measurements, was carried out with an HP 4156A Semiconductor Parameter Analyzer.

Flexible glass substrates with a thickness of 120 μm were sourced from Corning Incorporated. Voltage and current measurements were acquired using a National Instruments USB-6343 data acquisition (DAQ) card in conjunction with a ThorLabs AMP100 transimpedance amplifier. Layer thicknesses were measured using a KLA-Tencor P-7 profilometer. Substrate cleaning was performed using a reactive ion etching (RIE) system from Glow Research Inc. A BossLaser FM-D fiber laser desktop machine did the laser ablation.

Fabrication process

The fabrication process began with substrate cleaning to ensure optimal surface quality. Flexible glass substrates were cleaned via sonication for 10 min in acetone, followed by isopropanol (IPA), and subsequently treated using reactive ion etching (RIE) for 3 min to remove organic residues and activate the surface. A bilayer metal stack of chromium (Cr) and silver (Ag), with respective thicknesses of 5 nm and 30 nm, was deposited thermally to form the gate electrodes. Cr was deposited at a rate of 0.1 Å/s, followed by Ag at 1.5 Å/s, under a base pressure of 9 × 10−6 Torr.

A 20 nm layer of aluminum oxide (Al2O3) was then deposited using plasma-enhanced atomic layer deposition (ALD) at 200 ∘C, serving as a high-quality gate dielectric. An additional organic dielectric layer was deposited by chemical vapor deposition (CVD) of Parylene diX-SR, forming a 100 nm thick insulating film. The active organic semiconductors—S-DNTT-10 for p-type and TU-1 for n-type transistors—were thermally evaporated under a base pressure of 3 × 10−6 Torr. S-DNTT-10 was deposited at room temperature at a rate of 0.18 Å/s and subsequently annealed in ambient air at 60 ∘C. TU-1 was deposited at 70 ∘C at a rate of 0.1 Å/s after the source and drain deposition.

Access to the gate electrodes was established through the laser ablation of the dielectric layer. Source and drain contacts, as well as interconnect tracks, were fabricated by thermally depositing 30 nm of gold (Au) at 1.2 Å/s and 9 × 10−6 Torr. All circuit elements—including gates, contacts, semiconductor layers, and interconnects—were patterned using shadow masks to define the layout of the spiking neuron circuit. All experiments, including those demonstrating frequency modulation, coincidence detection, and adaptation, were performed within a 3-week period, during which device performance remained consistent with no observable degradation.