Flexible Neurons Made From Organic Materials Bring Brain Computing to Soft Robots

Low-power, bendable circuits process information like neurons, creating potential for adaptive robotics and bio-integrated devices.
A small chip bent between fingertips responds to light through its circuitry, adjusting LED output like a living retina. In contrast to the brittle silicon processors that power most computers today, this device flexes without breaking. It processes information through artificial neurons that fire and rest, modulating brightness based on ambient conditions. The system operates at speeds that match biological reflexes, consuming mere nanowatts of power.
This organic circuit reflects a growing effort in neuroscience-inspired computing, where materials and device designs aim to reproduce biological signal timing within deformable systems. As computing moves closer to biological models of perception and control, researchers are seeking materials that combine electrical responsiveness with mechanical flexibility. The capacity for circuits to bend and conform opens pathways for wearable health monitors, adaptive prosthetics, and robotic systems that integrate sensing and computation within soft, body-safe structures.
Scientists at Purdue University and collaborating institutions have created the first completely organic spiking neural circuit equipped with both excitatory and inhibitory connections. Published in npj Flexible Electronics, the work establishes a foundation for neuromorphic systems that can conform to curved surfaces, interface directly with living tissue, or provide computational intelligence within compliant robotic structures.
The system does not store long-term memories or learn persistent behaviors. Instead, it shows how spiking neural computation can occur within fully organic, mechanically flexible hardware.
The research team fabricated their devices on 120-micrometer flexible glass using shadow masking rather than photolithography. This approach kept the entire process outside cleanroom facilities while still achieving functional neural circuits. Each artificial neuron consists of complementary organic field-effect transistors (OFETs) paired with integrated capacitors formed from parasitic overlap between gate and channel electrodes.
Architecture Emulates Neural Computation
The circuit architecture follows the integrate-and-fire model, where incoming signals build until they reach a threshold that emits an output spike. The soma, or cell body equivalent, uses an Axon-Hillock design, a circuit configuration that reproduces the voltage threshold behavior seen in neurons, generating action potentials when membrane voltage crosses a critical level. Synaptic circuits employ pulsed current sources built from two OFETs operating in subthreshold regime, where transistors conduct minimal current with small gate voltages, producing exponentially controlled current pulses. Operating transistors this way allows the circuit to generate slow, graded currents similar to biological synapses without additional analog components.
Excitatory synapses use p-type semiconductors based on S-DNTT-10, while inhibitory connections rely on n-type TU-1 materials. The temporal dynamics of these synapses closely approximate biological systems, with excitatory connections reaching time constants near 65 milliseconds and inhibitory circuits extending to roughly 275 milliseconds, both within ranges observed in actual neural tissue where synaptic currents evolve over tens to hundreds of milliseconds.
The researchers characterized device behavior under both flat and mechanically bent conditions, applying tensile stress by wrapping substrates around a 205-millimeter radius cylinder. Bending produced opposing effects in the two transistor types. P-type devices exhibited threefold increases in carrier mobility and on-current, likely due to enhanced molecular ordering under strain. N-type devices showed twofold reductions, consistent with their greater sensitivity to structural disorder. Understanding how mechanical stress alters carrier mobility helps optimize circuit performance for flexible devices that must maintain consistent activity even when curved or stretched. Figures in the original study show the neuron’s response curves and flexibility tests, underscoring stable operation even when bent along this radius.
Computational Capabilities Across Multiple Scales
Testing revealed that the artificial neurons support several modes of neural computation. The firing rate response followed a sigmoidal curve, initiating spikes at 10 hertz with sub-nanoampere input currents and saturating near 55 hertz with 100 nanoamperes. This bounded output for unbounded input parallels the saturation behavior of biological neurons, where refractory periods limit maximum firing rates.
Linearity tests confirmed that when two excitatory synapses received simultaneous inputs, the output frequency matched the sum of individual responses. With each synapse producing roughly 0.6 hertz alone, concurrent activation yielded 1.25 to 1.27 hertz, confirming linear integration when inputs arrive with matched phases. This property simplifies network analysis and enables predictable signal summation across multiple synaptic connections.
Coincidence detection experiments showed temporal sensitivity within approximately 250 milliseconds. Two subthreshold inputs delivered with short intervals between them successfully triggered output spikes through temporal summation. When the inter-stimulus interval exceeded 260 milliseconds, the neuron failed to fire, indicating that synaptic currents had decayed below the integration threshold. This selectivity for temporally aligned inputs provides a mechanism for detecting synchronized events while filtering out asynchronous noise.
Real-Time Adaptation in Closed-Loop Control
Having confirmed the neuron’s core firing and integration performance, the researchers then explored whether it could support practical interaction within control systems. To validate operational functionality, they embedded their organic neuron in a feedback control system that regulated LED brightness based on ambient light measurements from a VEML 7700 sensor. In this testbed, the organic neuron served as the decision point linking input from a photometric sensor to output brightness in real time. An Arduino microcontroller read sensor values and adjusted both the presynaptic spike frequency and the synaptic weight voltage. Under dim conditions, the system reduced neural activity, dimming the LED. Bright ambient light increased firing rates and synaptic strength, raising LED output.
This short-term modulation differs from persistent learning mechanisms like long-term potentiation. The system adjusted synaptic parameters temporarily in response to environmental changes but did not encode lasting memories. The behavior nonetheless resembles biological short-term plasticity processes such as synaptic depression and facilitation, which enable sensory gating and dynamic filtering in nervous systems. Such reactive control could extend to robotic applications requiring continuous adjustment, from gripper force modulation to gait stabilization in legged robots. A soft robotic gripper, for example, could locally process tactile spikes within its elastomer skin, adjusting grip force before signals ever reach a central controller.
Power consumption remained below 10 nanowatts during baseline operation, with energy use scaling proportionally to firing frequency. Overall power use was governed mainly by somatic activity, with synaptic contributions an order of magnitude lower.
Material Properties Enable Biocompatibility
The organic semiconductors used avoid several problems associated with conducting polymers like PEDOT:PSS. Those materials rely on ionic transport and aqueous electrolytes, which compromise long-term stability and limit operating speed to frequencies below tens of hertz. PEDOT:PSS also exhibits acidic behavior and cytotoxicity that hinder direct biological interfaces.
The researchers instead employed electronic transport through OFETs, achieving response times suitable for spike frequencies up to 60 hertz. The fabrication stack included parylene as a dielectric layer and used aluminum oxide deposited via atomic layer deposition. Gold contacts formed source, drain, and interconnect pathways. All materials remained stable over the three-week testing period without observable degradation.
By avoiding ionic transport and using purely electronic conduction, these OFETs operate faster and remain stable without electrolyte degradation, expanding prospects for continuous operation in flexible sensors or soft prosthetic devices. Such material stability supports future designs for sensors or neural interfaces that must operate in wet or mobile environments over extended periods.
Device-to-device variability across eight positions on single chips showed threshold voltages varying within 150 millivolts for p-type and 120 millivolts for n-type transistors. Mobility spread across a 0.06 square centimeters per volt-second window. These variations remained small enough to support consistent circuit operation without extensive calibration.
Scaling Considerations and Future Directions
The following section examines how manufacturing methods and design choices affect the pathway from laboratory prototypes to practical neuromorphic systems. Current operating voltages ranged from 5 to 7 volts, optimized for circuit validation rather than biological safety. The underlying architecture supports voltage scaling for applications requiring direct tissue contact, though reducing operating voltage would require tighter control over threshold variability and leakage currents, which remain active areas of organic semiconductor research. Shadow mask resolution limited feature sizes and constrained maximum operating frequencies. Photolithographic patterning could reduce dimensions and push performance into the kilohertz regime, approaching native biological spiking rates.
The 14 by 5 millimeter footprint per neuron reflects the alignment tolerances of mask-based fabrication. Tighter control would enable higher neuron densities for more complex networks. The current chip contained 20 complete neurons plus 20 standalone transistors, showing manufacturability at modest scales.
Each neuron comprised six transistors: two p-type devices for excitatory synapses, two n-type for inhibitory connections, and two more for the soma. Parasitic capacitances between gate and source/drain regions provided the necessary charge storage. This approach eliminated discrete capacitor fabrication steps while maintaining temporal dynamics appropriate for neural computation.
Moving from vacuum deposition to solution-based processes could enable printed circuits on polymer films, making neuromorphic materials scalable and more accessible for integration with soft structures. While current devices used thermal evaporation for proof of concept, the selected materials and architecture remain compatible with solution-based methods using appropriate formulations. This path could reduce manufacturing costs and expand accessible substrate options.
Implications for Neuromorphic Engineering
The work fills a persistent gap in neuromorphic hardware development. Silicon-based processors excel at computational density and speed but remain incompatible with mechanical deformation or direct biological integration. Organic electrochemical devices offer flexibility and ionic conductivity but suffer from slow response times and stability concerns. Organic field-effect implementations occupy an intermediate position, providing electronic conduction speeds with mechanical compliance.
The integration of both excitatory and inhibitory synapses within a single flexible system enables implementation of balanced networks, where opposing inputs shape activity patterns. Without inhibitory connections, spiking networks lose control over timing and gain, often becoming unstable. Including both excitatory and inhibitory synapses allows activity to remain bounded and temporally precise. Inhibition serves critical functions in biological brains, including gain control, temporal precision, and pattern separation. Artificial systems lacking inhibitory components forfeit these capabilities.
Time constant tunability through synaptic weight voltages allows adaptation to different computational operations. Fast excitatory dynamics support rapid responses, while slower inhibitory currents enable sustained modulation. This combination provides temporal filtering across multiple scales within individual neurons.
The closed-loop light control validation showed that organic neurons can participate in sensorimotor systems requiring continuous environmental interaction. Such applications demand real-time processing rather than batch computation on stored datasets. The ability to modulate behavior based on immediate sensory feedback makes these devices suitable for autonomous robots, adaptive prosthetics, or implantable neural interfaces.
Future work could explore learning mechanisms that produce persistent weight changes. Spike-timing-dependent plasticity, where synaptic strength adjusts based on the relative timing of pre- and postsynaptic spikes, has been implemented in various neuromorphic substrates. Organic implementations could combine this capability with mechanical flexibility for adaptive systems that conform to complex surfaces.
Bridging Computation and Physical Form
Conventional computing architectures separate processing from sensing and actuation through discrete modules connected by wired interfaces. Neuromorphic systems embedded within structural materials could distribute intelligence throughout a system rather than concentrating it in a central processor. A robot limb might contain neural circuits within its compliant structure, processing tactile feedback locally before transmitting only high-level signals to other regions.
Biological organisms achieve this integration naturally, with neurons residing within soft tissue and connecting through flexible axons and dendrites. The organic spiking neuron constitutes progress toward artificial systems with comparable physical properties. The combination of mechanical flexibility, low operating power, and biologically relevant temporal dynamics makes these devices suitable for applications where rigid silicon remains inappropriate.
The research extends earlier work by the same group on analog neuromorphic circuits and synaptic components. Previous implementations included organic memristors and non-spiking neural primitives. The incorporation of spiking dynamics adds temporal coding capabilities absent from purely analog implementations. Spike timing carries information through the precise moments when neurons fire rather than through continuously varying voltages. This coding scheme supports coincidence detection and other computations difficult to implement with rate-based signals alone.
The team also previously developed the NUCLEOs system, a hybrid organic-inorganic neuromorphic robot capable of legged locomotion. That work combined organic sensors and actuators with conventional control electronics. The current realization of fully organic neurons suggests pathways toward entirely soft, self-contained neuromorphic systems that eliminate rigid components entirely.
Future prototypes might integrate sensory feedback and movement control in a single material layer, creating robotic surfaces capable of local decision-making and coordinated response without external processors. By merging electronic performance with pliable materials, this work points toward devices capable of continuous adaptation within biological or robotic environments, a link between computation and physical form that continues to evolve.
Source: Mirshojaeian Hosseini, M.J., Yang, Y., Bamford, S. et al. An organic spiking artificial neuron with excitatory and inhibitory synapses: towards soft and flexible organic neuromorphic processing. npj Flexible Electronics (2026). https://doi.org/10.1038/s41528-025-00512-6
