Recent advances in a-IGZO synaptic transistors for flexible computing applications

Working mechanisms of flexible synaptic transistors

The ability of biological neural networks to process temporal information arises in part from activity-dependent changes in synaptic efficacy. As shown in Figure 2, synaptic transmission involves neurotransmitter release, receptor activation, and ion-channel gating^[24]. Depending on the strength, timing, and repetition of neural activity, synapses can exhibit transient responses, referred to as short-term plasticity (STP), or longer-lasting changes associated with long-term plasticity (LTP)^[25,26]. These forms of plasticity contribute to temporal filtering, learning, and memory. Artificial synaptic devices do not reproduce the underlying biological chemistry directly; instead, they map electrical or optical input pulses onto conductance changes with comparable temporal characteristics.

Figure 2. Biological-to-artificial synaptic mapping. Schematic illustration showing the structural and functional analogy between a biological synapse and a flexible artificial synaptic transistor, mapping neurotransmitter dynamics to channel conductance modulation. Adapted from Ref.^[24] under the CC BY 3.0 license. Copyright 2015, The Authors. EPSC: Excitatory postsynaptic current; a-IGZO: amorphous indium gallium zinc oxide.

Among the candidate device structures, three-terminal a-IGZO TFTs provide a convenient platform for this functional mapping. The gate applies the presynaptic stimulus, the channel current represents the postsynaptic response, and the gate dielectric or electrolyte mediates conductance modulation. The suitability of a-IGZO as the channel material is associated with its disorder-tolerant electronic transport. The CBM of a-IGZO is mainly formed by the overlap of spatially extended and nearly isotropic 5s orbitals of heavy-metal cations, particularly In^3+[15]. Unlike the strongly directional sp³ orbitals of covalently bonded semiconductors such as Si and GaAs, these orbitals are less sensitive to variations in bond angles. Effective wavefunction overlap and electronic percolation can therefore be maintained despite the absence of long-range atomic order^[11,17,18].

This disorder-tolerant transport mechanism enables electron mobility above 10 cm²/Vs in optimized a-IGZO films and avoids the grain-boundary-related transport variations found in polycrystalline semiconductors. This mobility is typically one to two orders of magnitude higher than that of hydrogenated amorphous silicon (a-Si:H). Stable electrical operation has also been reported during bending in appropriately designed flexible metal-oxide TFTs^[9]. For example, a-IGZO TFTs using IZO/Ag/IZO source/drain electrodes exhibited only a 5.37% reduction in field-effect mobility and a threshold-voltage shift from 0.21 to 0.72 V after 10,000 bending cycles at a radius of 10 mm^[11]. However, isotropic orbital transport does not itself prevent mechanical cracking. The bending reliability of a complete device also depends on the thickness and position of the active layer, the mechanical properties of the dielectric and electrodes, interfacial adhesion, and strain distribution.

To emulate STP, many a-IGZO synaptic transistors use electric-double-layer (EDL) coupling at the electrolyte/channel interface. A gate pulse drives mobile ions, such as H⁺ or Li⁺, toward the interface, where they induce a high density of image charges in the channel. Their subsequent back-diffusion produces a transient conductance response and time-dependent relaxation^[27]. The large interfacial capacitance allows substantial channel modulation at relatively low gate voltages, while the relaxation time depends on ion mobility, electrolyte composition, and pulse conditions. Longer-lasting conductance states require less volatile storage processes. Charges can be retained in deep traps within the dielectric or at the interface, in oxygen-vacancy-related states, or in an embedded floating gate, thereby shifting the threshold voltage and preserving the programmed conductance^[28]. The resulting retention depends on trap depth, interfacial barriers, and the kinetics of charge release.

Ion migration and charge storage provide electrically driven plasticity but do not directly couple optical sensing to synaptic-weight modulation. This motivates optoelectronic a-IGZO devices, in which photoexcitation and persistent photoconductivity (PPC) allow light to act as both a sensory input and a programming stimulus. Illumination generates carriers and changes the occupancy of oxygen-vacancy-related states; carrier trapping or slow recombination can then sustain an elevated channel conductance after illumination^[29,30]. Thus, optoelectronic a-IGZO transistors can couple optical sensing with conductance modulation, although the magnitude and persistence of the response depend strongly on defect states, illumination conditions, and device structure.

Performance evaluation of flexible synaptic transistors

Constructing flexible neuromorphic systems capable of conformally interfacing with biological tissues while performing edge computing requires a balance between the physical limitations of devices and the functional requirements of bio-inspired computing. In practical applications, such as electronic skins attached to joints, devices often operate under mechanical deformation. Flexible synaptic transistors must therefore combine mechanical deformability with reliable electrical computation, as neither property alone is sufficient for functional computing systems. Accordingly, mechanical compliance and electrical consistency should be evaluated together. Existing evaluation frameworks do not fully capture the coupled electrical and mechanical characteristics of flexible synaptic transistors^[5]. This section therefore establishes a multidimensional evaluation framework covering algorithmic fidelity, energy efficiency and stability, and mechanical robustness. It evaluates both the computational accuracy of these devices as artificial synapses and their operational reliability as flexible electronic components under complex mechanical stresses.

(I) Linearity and symmetry are important metrics for assessing synaptic devices used in on-chip training. In the resistive processing unit (RPU) model proposed by Gokmen and Vlasov, an ideal weight update requires the conductance increment (ΔG) produced by each programming pulse to remain independent of the current conductance state^[31]. In a-IGZO devices, nonlinear and asymmetric conductance updates can arise from nonuniform barriers during charge trapping and detrapping or from the accumulation of mobile ions at the interface. These deviations introduce weight-update errors and may reduce network convergence and training accuracy^[32]. D2D variability introduces additional errors in large arrays, making fabrication uniformity as important as the intrinsic update characteristics. In an all-photolithographic array reported by our group, controlled device geometry and a homogeneous composite-electrolyte response narrowed the device distribution and reduced accumulated weight-update errors. Device-based simulations achieved a handwriting-recognition accuracy of 92.8%, with an interdevice standard deviation of 9.69%^[22]. Because the effect of nonlinearity also depends on the programming scheme and network model, these device metrics should be reported together with the pulse conditions and evaluation method.

(II) Dynamic range (G_max/G_min) and conductance-state resolution are complementary metrics that determine how precisely synaptic weights can be represented. Although biological synapses exhibit analog changes in efficacy, hardware neural networks require a sufficient number of distinguishable and reproducible conductance states. More than 100 states correspond to a resolution exceeding 6 bits, although the required resolution depends on the network architecture and training method. The number of nominal states alone is insufficient; state-to-state variation and read noise determine whether adjacent conductance levels remain distinguishable. In flexible devices, polymer dielectrics and low-temperature-processed gate stacks may provide insufficient gate capacitance or increased leakage, thereby restricting the accessible dynamic range. Channel and gate-geometry engineering, including superlattice channels and coplanar gates, has been used to broaden the operating range without exceeding the thermal budget of flexible substrates^[33,34].

(III) Efficiency and stability should be evaluated together because low-voltage operation does not necessarily ensure long-term conductance stability. For comparison, the energy associated with a biological synaptic event is commonly estimated to be approximately 1-10 fJ^[35]. In electrolyte-gated transistors (EGTs), EDL capacitances above 1 μF·cm^-2 can support operating voltages below 0.5 V. Attojoule-level energy consumption per event has been reported in synaptic transistors incorporating electrospun nanofiber interfaces^[22,36]. However, spontaneous ion diffusion can limit the retention of the programmed state. Composite electrolytes and three-dimensional interfaces have therefore been explored to regulate ion-migration pathways and barriers, providing a means to balance programming energy against retention time^[37,38]. Reported energy values should specify the pulse amplitude, pulse width, current-integration method, and device area, while retention should be measured under defined bias and environmental conditions.

(IV) Critical bending radius (R_c) is a common static measure of mechanical flexibility. Unlike rigid devices, flexible a-IGZO transistors are exposed to tensile and compressive strains during operation. According to neutral-plane mechanics, excessive strain in brittle inorganic layers can initiate microcracks, particularly when these layers are positioned away from the neutral mechanical plane^[39]. Interfacial delamination may introduce an additional failure pathway. Compliant organic electrolytes, including polyvinyl alcohol (PVA)/lignin composites, can reduce elastic-modulus mismatch and suppress crack propagation, thereby improving tolerance to bending or folding^[38]. Nevertheless, R_c should be reported together with device thickness, bending direction, applied strain, and the electrical state during testing because bending radius alone does not fully describe mechanical reliability.

(V) Fatigue life evaluates reliability under repeated bending or stretching and is more relevant to long-term wearable operation than a single static bending test. During more than 10⁴ loading cycles, accumulated structural defects may increase contact resistance or eventually produce open circuits. Viscoelastic substrates and compliant device layers can dissipate strain energy and delay damage accumulation^[40]. A small initial R_c therefore does not necessarily indicate high resistance to cyclic fatigue. Electrical characteristics and synaptic responses should be measured both during and after cyclic loading.

(VI) Electromechanical stability under dynamic deformation describes whether synaptic operation remains reproducible during deformation. Ideally, key characteristics - including spike-timing-dependent plasticity (STDP) time windows, pulse-response amplitudes, conductance-update trajectories, and retention - should remain within defined tolerances under different strain states. In practice, mechanical deformation can perturb ion transport, interfacial contact, and charge modulation, causing identical electrical stimuli to produce different synaptic responses, as observed in stretchable synaptic transistors^[40]. Such strain-induced variations may reduce the inference accuracy of a network trained under undeformed conditions. Consequently, changes in synaptic parameters should be quantified as a function of strain and loading cycles rather than inferred only from electrical measurements performed before and after bending.

Together, these metrics highlight the need to balance electrical performance with mechanical reliability in flexible a-IGZO synaptic transistors. These trade-offs must also be considered alongside the fabrication challenges associated with a-IGZO TFT technologies^[41]. A representative flexible a-IGZO synaptic transistor has demonstrated femtojoule-level energy consumption and a widely tunable memory time^[42]. Nevertheless, device linearity, retention, and response consistency remain strongly dependent on the dielectric, interfaces, and deformation conditions. The following sections discuss material- and device-architecture strategies for improving these properties while preserving mechanical durability.

Materials engineering

Channel-material selection establishes the fundamental trade-offs between electrical performance and mechanical compliance in flexible neuromorphic devices. Representative conjugated polymers and a-IGZO occupy complementary positions in this respect. As summarized in Table 2, a-IGZO generally provides higher carrier mobility and better electrical stability, whereas conjugated polymers offer greater intrinsic softness and tolerance to tensile strain.

For a-IGZO, materials engineering is therefore central to preserving its favorable carrier transport and electrical stability while improving energy efficiency, functional response, and compatibility with the processing and operating requirements of flexible neuromorphic devices. Because a single compositional adjustment rarely provides high mobility, reliable conductance modulation, and multimodal sensing simultaneously, recent studies have combined several forms of material control. Rather than relying only on deposition optimization, these approaches target the electronic structure, ionic dynamics, and interfacial energy landscape that govern carrier generation, transport, and storage. They include modifying channel composition and defects, functionalizing the back channel, regulating ion motion or ferroelectric polarization in the gate dielectric, and introducing heterointerfaces to modify optical absorption and carrier separation. The following subsections examine these strategies together with the processing and stability constraints relevant to flexible devices.

Regulation of channel atomic components

In amorphous oxide semiconductors (AOSs), metastable oxygen-related defects, particularly V_o, play a major role in threshold-voltage shifts (ΔV_th) under bias stress and can also contribute to nonlinear conductance updates in synaptic devices. In the archetypal ternary In-Ga-Zn-O system, the electronic properties of the film are governed by the orbital characteristics and metal–oxygen bonding of its constituent cations. The spatially extended, nearly isotropic In 5s orbitals provide the principal electron-transport pathways, whereas the stronger affinity of Ga for oxygen suppresses excessive carrier generation and oxygen-vacancy formation; Zn contributes to the formation of the amorphous oxide network^[17,18,52]. The stabilizing effect of Ga is commonly associated with the higher reported bond dissociation energy of Ga-O (~374 kJ/mol) relative to In-O (~346 kJ/mol)^[52]. Consequently, the comparatively weaker In-O bonding may make In-rich compositions more susceptible to oxygen loss and defect formation during energetic processing or prolonged electrical or optical stress. Increasing the Ga content can improve stability, but the smaller spatial extent of Ga 4s orbitals may impede carrier transport. This composition-property relationship also explains why nominal stoichiometry cannot be evaluated independently of deposition and post-treatment conditions. Channel composition must therefore balance carrier mobility against defect-related stability.

One approach to this trade-off is to introduce an additional metal cation that modifies the electronic states near the CBM. The heavy-metal-cation orbital-overlap model proposed by Hosono identifies the overlap of spatially extended ns orbitals as an important factor governing electron transport in amorphous oxides^[15]. On this basis, Sn can replace Ga to form indium–tin–zinc oxide (ITZO) or partially substitute for Ga to form indium–gallium–zinc–tin oxide (IGZTO)^[53]. Because Sn and In are both fifth-period cations with spatially extended 5s orbitals, Sn provides a transport contribution more similar to that of In than to the smaller Ga 4s orbital. Overlap between the Sn and In 5s states can form broader and more continuous electronic percolation pathways near the CBM. This modification can reduce the electron effective mass and increase field-effect mobility while retaining some suppression of oxygen-vacancy formation^[54]. The wider percolation network may also support rapid modulation of channel conductance during synaptic operation. However, the resulting carrier concentration and subthreshold characteristics remain sensitive to composition, oxygen partial pressure, and other deposition conditions^[53,54].

When bias stability is prioritized, Hf-based channel engineering provides a complementary strategy for regulating oxygen-related defects. An Hf-doped a-IGZO back-channel layer can act as an oxygen scavenger and redistribute oxygen near the channel interface, while Hf-doped IGZO layers can suppress V_o and other oxygen-related defects, thereby improving bias stability^[55,56]. However, the effects of Hf remain concentration-dependent, because excessive suppression of oxygen-related donors can reduce carrier density and alter mobility. Location-selective chloride modulation doping provides another means of controlling defect states, as demonstrated in the dual-channel device shown in Figure 3A(i)^[57]. The ion-regulated band model in Figure 3A(ii) illustrates how chloride incorporation modulates the electronic structure of the doped channel. Chloride incorporation suppresses persistent photocarrier excitation and reduces the density of oxygen-vacancy-related subgap states, consistent with the O 1s X-ray photoelectron spectroscopy results shown in Figure 3A(iii). The electrical measurements further indicate reduced threshold-voltage drift in the doped channel^[57]. Unlike uniform channel doping, this spatially selective approach can improve stability without modifying the entire transport channel.

Figure 3. Materials engineering and microscopic mechanism regulation of a-IGZO synaptic transistors. (A) Channel atomic components: (i) Schematic illustration of the location-selective chloride (Cl^-) modulation-doped IGZO dual-channel synaptic transistors. (ii) Energy band and ion distribution model during the potentiation process, elucidating the ion-regulated synaptic weight update mechanism. (iii) O 1s XPS spectra and XRD patterns of the channel regions, demonstrating the passivation and modulation of V_o levels via Cl-doping. Adapted from Ref.^[57] under the CC BY 4.0 license. Copyright 2025, The Authors; (B) Back-channel engineering: (i) 3D schematic illustration of an IGZO phototransistor featuring an ASL at the back-channel, alongside the cross-sectional image and EDS mapping highlighting the oxygen-deficient interface. (ii) Proposed subgap DOS models for the IGZO phototransistors without and with the ASL, illustrating the widened subgap defect levels. Adapted from Ref.^[62] under the CC BY 4.0 license. Copyright 2025, The Authors; (C) Molecular dynamics of dielectric: (i) Schematic of an individual EGT in a high-density artificial synapse array (100 × 100) with a lateral-gate structure, highlighting the ionic conduction framework enhanced by ZIF-67 porosity. (ii) Representative PPF and long-term EPSC responses, showing the enhanced synaptic performance. Adapted from Ref.^[22] under the CC BY 4.0 license. Copyright 2023, The Authors; (D) Heterojunction bandgap engineering: (i) Schematic of emulating a biological synapse using an IGZO/CsPbBr₃ NPs/IGZO TFT architecture under electrical and optical stimuli. (ii) Schematic energy-band diagrams of the IGZO/CsPbBr₃ NPs/IGZO heterostructure in the dark and under illumination, illustrating photocarrier separation and transfer. Adapted with permission^[76]. Copyright 2021, American Chemical Society. a-IGZO: Amorphous indium gallium zinc oxide; XPS: X-ray photoelectron spectroscopy; XRD: X-ray diffraction; ASL: aluminum sensitization layer; EDS: energy-dispersive X-ray spectroscopy; DOS: density of states; EGT: electrolyte-gated transistor; PPF: paired-pulse facilitation; EPSC: excitatory postsynaptic current; NPs: nanoparticles; TFT: thin-film transistor; ITO: indium tin oxide; PEO: poly(ethylene oxide); PVK: perovskite.

Post-treatment provides an additional route for passivating dangling bonds and nanovoids produced during low-temperature film formation. Hydrogen annealing can passivate some defects, although excessive hydrogen may introduce donor states, increase channel conductivity, and degrade switching behavior. Nitrogen-plasma treatment has also been used to passivate oxygen-vacancy-related sites through the formation of metal–nitrogen bonds. In an IGZO-based resistive memory device, this treatment improved resistive-switching stability and carrier injection at the contact region^[58]. This result supports the underlying passivation mechanism, although the reported device is not a three-terminal synaptic transistor.

Supercritical-fluid (SCF) treatment provides a low-temperature method for transporting reactive molecules into dense oxide films. Unlike conventional gas-phase annealing, which may be limited by diffusion into dense films, a supercritical medium combines gas-like diffusivity with liquid-like solvent capacity. Chen et al. used supercritical CO₂ at temperatures above 31 °C and pressures above 7.38 MPa to deliver polar H₂O or NH₃ molecules into the film^[20]. These properties facilitate oxidation and hydrogen-bond-network reconstruction, reducing deep subgap defects and improving synaptic-weight linearity. Density functional theory (DFT) calculations further indicate that moderate nitrogen incorporation and the resulting metal–nitrogen bonds can modify the valence-band maximum (VBM), suppress unwanted hole-related processes, and improve light stability and current modulation^[59]. For large-area flexible devices, the uniformity of molecular delivery and passivation across the channel remains an important processing consideration. Overall, compositional doping and low-temperature passivation provide complementary means of balancing mobility and defect stability, but their effectiveness depends strongly on dopant concentration, spatial distribution, and process control; spatially selective treatments may therefore be particularly suitable for flexible devices with restricted thermal budgets.

Structural reconfiguration

Material engineering modifies charge transport and storage at the material and interface levels, whereas structural engineering controls gate coupling, current pathways, and integration density at the device level. Conventional planar single-gate architectures may face trade-offs among linear weight modulation, operating voltage, and cell density. Recent studies have therefore explored several complementary strategies, including gate-stack design, 1D nanochannels, vertical integration, and multi-terminal architectures. Representative architecture-engineering strategies are illustrated in Figure 4 and discussed in detail in the following subsections.

Figure 4. Structural reconfiguration of flexible a-IGZO synaptic transistors. (A) Floating-gate stack engineering: (i) Schematics of IGZO synaptic transistors incorporating IGZO and ITO floating gates, respectively. (ii) Transfer characteristics of the two devices measured under a gate double sweep from -15 to 15 V. (iii) Retention characteristics of the programmed and erased states for devices with IGZO and ITO floating gates. Adapted from Ref.^[28] under the CC BY 4.0 license. Copyright 2025, The Authors; (B) 1D Nanochannel architecture: (i) 3D schematic of an IGZO/In₂O₃ NW heterojunction phototransistor stimulated by a UV light pulse. (ii) Transient current responses at various pulse frequencies and the corresponding extracted SRDP. Adapted with permission from^[89]. Copyright 2024, Elsevier; (C) Vertical 3D integration: (i) Cross-sectional schematic and TEM image of a vertical-channel synapse TFT, minimizing device footprint via HfO₂ spacer engineering. Adapted with permission^[91]. Copyright 2025, American Chemical Society. (ii) Schematic of high-performance IGZO CAA transistors featuring in situ engineered vertical indium gradients. (iii) Comparison of vertical In:Ga atomic ratios at deposition pressures of 30 mTorr. Adapted with permission^[92]. Copyright 2026, American Chemical Society; (D) Multi-terminal collaborative architecture: (i) Illustration of a hetero-synapse excited by electric inputs from two presynaptic terminals and the corresponding electrochemical ion modulation of the channel layer under dual-terminal gating. (ii) Postsynaptic currents (EPSCs) excited by five presynaptic spikes separately or simultaneously applied to G1 and G2 terminals. (iii) Architecture of the neural network based on the hetero-synaptic transistor for MNIST image recognition. (iv) Comparison of classification accuracy between networks based on the dual-gate hetero-synaptic transistor and the single-gate homo-synaptic transistor. Adapted with permission^[105]. Copyright 2022, The Royal Society of Chemistry. a-IGZO: Amorphous indium gallium zinc oxide; ITO: indium tin oxide; NW: nanowire; UV: ultraviolet; SRDP: spike-rate-dependent plasticity; TEM: transmission electron microscopy; TFT: thin-film transistor; CAA: channel-all-around; EPSCs: excitatory postsynaptic currents; MNIST: Modified National Institute of Standards and Technology; PL: protection layer; CH: channel; TL: tunneling layer; FG: floating gate; PAD: pad (contact pad); GI: gate insulator; ACT: active layer.

In-sensor computing

In-sensor computing performs signal conditioning or feature extraction within the sensing element or sensor array. By processing analog signals before transmission, this architecture can reduce back-end data conversion and communication requirements, although it does not eliminate system latency or peripheral-circuit overhead^[114].

In a-IGZO optoelectronic synaptic transistors, PPC provides one mechanism for coupling optical sensing with conductance memory. Under ultraviolet or short-wavelength visible illumination, photoexcited electrons contribute to channel current, while holes may be trapped at defect-related states in the bulk or at interfaces. The resulting separation of carriers can delay recombination after illumination is removed, producing a conductance response that depends on pulse intensity, duration, and repetition^[115,116]. This response can support photodetection, temporal integration, and the emulation of synaptic plasticity within the same device, but its retention and recovery are sensitive to defect density and environmental conditions.

To break through a-IGZO’s intrinsic bandgap limitations and extend its spectral response into the visible region, bandgap engineering and heterojunction strategies are essential. For instance, Jo et al. demonstrated that blending dyes with PMMA to form heterojunctions with a-IGZO phototransistors significantly enhances visible-light responsivity^[117]. Related ZnO/MoS₂ memristors have demonstrated stable optoelectronic switching and have been used for recognition tasks^[118]. Because these devices do not employ a-IGZO, they are included only as a comparative example of heterojunction-enabled optoelectronic memory. In parallel, Pereira et al. employed hydrogen doping to enhance the visible-light sensitivity of IGZO optoelectronic memristors, enabling synaptic responses to blue and green light^[119].

Beyond spectral broadening, optimizing carrier separation and absorption efficiency is equally critical. The a-IGZO/InHfO_x nanowire device designed by Luo et al. utilized the 1D confinement effect to significantly reduce energy consumption while achieving an ultra-high dynamic range of 74.9 dB^[90]. Meanwhile, to meet the extreme demand for weak light detection, Kwak et al. modified the a-IGZO surface with an ultrathin aluminum (Al) nanofilm to enhance the absorption cross-section via the LSPR effect. The constructed optical sensor array maintained high-SNR image perception capability under an extremely low light intensity of 10 μW/cm^2[62]. To further ensure the computational precision of these analog arrays, Seo et al. adopted in situ defect engineering via precisely controlling V_o to passivate uncontrolled surface traps. This highly linear conductance regulation allowed the simulated neural network to achieve an inference accuracy of 92% on the Modified National Institute of Standards and Technology (MNIST) task^[120].

Moving beyond static intensity perception and simple pattern recognition, processing dynamic spatiotemporal visual information directly within the sensory array represents a higher-level goal for in-sensor computing. Leveraging the PPC and tunable charge trapping dynamics of a-IGZO, optoelectronic synapses can naturally integrate spatial and temporal information. For instance, Mondal et al. developed an a-IGZO-based transparent optoelectronic synapse that natively encodes the temporal sequence of moving light stimuli into continuous conductance states, realizing stealth mode trajectory tracking and registration directly at the hardware level^[121]. Furthermore, to cope with complex and changing illumination environments, simulating the dynamic adaptation of the biological retina is equally crucial. By synergistically modulating the optical stimulus and electrical gate bias, a-IGZO synaptic transistors can dynamically shift their photo-responsivity, effectively emulating the transition between scotopic (dark-adapted) and photopic (light-adapted) vision. This environment-adaptive plasticity enables the artificial visual system to automatically prevent overexposure under strong illumination while maintaining high sensitivity in dim conditions, significantly expanding the dynamic range of flexible machine vision without the need for redundant back-end algorithms^[87].

In the field of large-scale flexible integrated manufacturing, all-inkjet printing offers a low-cost pathway for scalable bio-inspired arrays. As a pioneering example in the broader family of AOSs, Liang et al. reported printed optoelectronic synapses based on SnO₂/InP/ZnSe quantum dot heterostructures, achieving crucial synaptic behaviors with an ultralow power consumption of ~5.6 pJ per event and 91% recognition accuracy^[122]. Extending this methodology to tactile sensing, Li et al. developed a neuromorphic pressure afferent nerve that exhibited a rapid threshold response (< 4 s) and 99.3% accuracy in spatially localized pressure mapping^[123]. Furthermore, Franco et al. verified the versatility of such inkjet-printed IGZO devices in achieving reliable volatile and nonvolatile switching modes^[124], highlighting the immense potential of printing technologies for scalable sensory prostheses.

As neuromorphic hardware evolves, advancing from single-sensory perception toward multimodal fusion and system-level adaptation becomes crucial. To this end, Hou et al. explored photonic–electronic synergistic modulation for energy-efficient multimodal learning^[125]. Wen et al. further demonstrated dynamic visual–thermal multimodal perception based on defect-modulated electrospun nanofibers^[126]. For the secure transmission of these multimodal signals, Yang et al. fine-tuned the 1/f noise characteristics of ultra-thin ALD-driven transistors via interface defect engineering, achieving stable encrypted opto-acoustic perception^[127]. Alongside these device-level functions, array-level demonstrations are needed to determine whether sensory information can be processed directly in physical a-IGZO hardware. Kang et al. developed optoelectronic synaptic memtransistors based on a photoresponsive IGZO channel and an HfO₂ contact-engineered structure. Inspired by the organization of photoreceptors in the biological retina [Figure 5A(i)], the authors fabricated a 6 × 6 optoelectronic synaptic memtransistor (OSMT) array in which each device represented one image pixel [Figure 5A(ii)]. Successive electrical or optical write–erase operations increased the current difference between the target pattern and its background, thereby enhancing the contrast of the mapped image [Figure 5A(iii)]. This work demonstrates optical sensing, conductance modulation, and elementary image preprocessing within a physical IGZO device array, although larger arrays and flexible implementations remain to be demonstrated^[128].

Figure 5. Integrated sensing-storage-computing systems and application scenarios. (A) Retina-inspired in-sensor image processing using contact-engineered IGZO optoelectronic synaptic memtransistors: (i) schematic of the biological retina and its photoreceptor arrangement; (ii) optical micrograph of the OSMT array, the corresponding 6 × 6 device array, and its initial current mapping; and (iii) evolution of the mapped image under successive electrical and optical write–erase operations, showing increased contrast between the target pattern and its background. Adapted from Ref.^[128] under the CC BY 4.0 license. Copyright 2026, The Authors; (B) Tactile perception and robotic navigation based on FST encoding: (i) Schematic of the FST tactile encoding system. (ii) Real-time classification results for autonomous vehicle status detection and smart robotic navigation using the PDTFT-based SNN hardware. Adapted from Ref.^[136] under the CC BY 4.0 license. Copyright 2025, The Authors; (C) RC for handwritten digit recognition: (i) Architecture of the RC system using IGZO-based leaky 2T dynamic random-access memory (DRAM) nodes as the physical reservoir. (ii) Confusion matrix showing the recognition results for the MNIST pattern recognition task. Reprinted with permission^[141]. Copyright 2024, American Chemical Society; (D) Analog CIM and on-chip training: (i) Circuit topology of the 6T1C synaptic cell consisting of n-type IGZO TFTs and a storage capacitor, alongside a photograph of the fabricated 8-inch wafer. (ii) Synaptic weight update curves (potentiation and depression) and memory retention characteristics of the single 6T1C device. Adapted from Ref.^[147] under the CC BY 4.0 license. Copyright 2023, The Authors. IGZO: Indium gallium zinc oxide; OSMT: optoelectronic synaptic memtransistor; FST: first-spike-timing; PDTFT: pressure-sensitive dual-TFT; SNN: spiking neural network; RC: reservoir computing; 2T: two-transistor; DRAM: dynamic random-access memory; MNIST: Modified National Institute of Standards and Technology; CIM: in-memory computing; TFTs: thin-film transistors; LECTS: light-electric coupling synaptic transistor; EPSC: excitatory postsynaptic current; PPF: paired-pulse facilitation; WLU: word line upper; BLU: bit line upper; GND: ground; WLD: word line lower; BLD: bit line lower; ADC: analog-to-digital conversion.

Li et al. further demonstrated an a-IGZO-based optoelectronic nociceptive device that generated nonlinear responses to increasing stimulus intensity and was coupled to an end-effector retraction operation with a reported response time below 10 ms^[81]. This result illustrates the potential of sensor-level threshold responses for reducing control latency, although the overall response also depends on the associated readout and actuation circuitry. In-sensor a-IGZO devices can therefore combine optical sensing with temporal conductance modulation, but their broader use remains limited by spectral range, recovery dynamics, device variability, and system-level integration.

Near-sensor computing

Near-sensor computing places signal-processing circuits adjacent to the sensor rather than within the sensing element itself. In flexible health-monitoring systems, this arrangement can perform amplification, filtering, event detection, or feature extraction before wireless transmission, thereby reducing the amount of raw physiological data sent to a remote processor. Its effectiveness depends on the power consumption, mechanical compatibility, and parameter uniformity of both the sensor and the local processing circuits.

Human physiological signals [such as electroencephalography (EEG) and electromyography (EMG)] are typically characterized by weak amplitudes (ranging from microvolts to millivolts, e.g., 10 μV to 5 mV)^[129], low frequencies, and susceptibility to motion interference. Because conventional rigid readout circuits cannot conform readily to skin deformation, the large-area uniformity and mechanical compatibility of a-IGZO TFTs make them promising building blocks for flexible front-end amplification circuits^[130]. Particularly for analog circuits such as differential amplifiers, the consistency of device parameters (such as threshold voltage and mobility μ) is crucial. The all-photolithographic integration process recently developed by Li et al. successfully controlled the deviation of device parameters within the array to less than 5%, providing a solid hardware foundation for high-performance flexible analog front-end circuits^[22]. On this basis, Ishida et al. designed a pseudo-complementary metal-oxide-semiconductor (CMOS)-based operational amplifier with a-IGZO TFTs on a flexible film, successfully achieving a 22.5 dB open-loop gain and a 31 kHz gain-bandwidth product, enabling weak bioelectric signals to be amplified directly to levels recognizable by back-end digital logic^[131]. Zhao et al. further developed a unipolar a-IGZO TFT amplifier circuit with a parallel bootstrap capacitor, achieving an exceptional gain of 43.5 dB for bio-signal sensing applications, demonstrating excellent robustness for physiological monitoring^[132].

Beyond simple signal amplification, near-sensor computing modules must also undertake feature extraction tasks. Shim et al. developed an all-rubber-based “cognitive skin”, integrating stretchable a-IGZO synaptic transistor arrays directly at the back-end of multi-channel tactile sensors. The system cleverly utilizes the exponential current-voltage characteristics of a-IGZO devices in the subthreshold region to construct high-pass filters (cutoff frequency < 1 Hz), effectively filtering out baseline drift noise caused by slow skin deformation^[19]. Yuan et al. further constructed an artificial tactile perception system with spatio-temporal recognition capabilities, utilizing the STP of synaptic transistors to encode temporal signals generated by sliding, achieving material and pattern classification directly at the edge^[133]. By natively decoding these complex spatio-temporal tactile patterns locally, the near-sensor computing module substantially reduces the volume of raw sensory data that needs to be transmitted to central processors, providing an energy-efficient approach for flexible electronic skin.

To compress data transmission volume to the extreme, near-sensor computing modules based on a-IGZO have further introduced the “event-driven” mechanism of the biological nervous system. In this architecture, synaptic devices are configured as integrate-and-fire (IF) neuron models, sending pulse signals to the back-end only when the cumulative intensity of the input signal exceeds a set threshold. Utilizing the unipolar switching characteristics of a-IGZO transistors, Lebanov et al. developed flexible IF neuron circuits for healthcare applications^[134]. These circuits accumulate input signals and generate an output spike only after the membrane-like voltage exceeds a defined threshold, allowing infrequent or high-amplitude events to be encoded into sparse pulses. This operating principle is relevant to anomaly-triggered physiological monitoring, although its effect on complete-system battery life requires evaluation together with the sensor, wireless transmitter, and peripheral circuits.

The advantages of near-sensor computing also extend to complex spatiotemporal feature extraction, reducing reliance on energy-intensive ADCs. To achieve seamless continuous-to-discrete signal conversion, Qiao et al. constructed 1T1M-based universal oscillating units to serve as artificial sensory neurons, dynamically translating analog stimuli into highly compressed discrete spike trains^[135]. Extending neuromorphic perception toward tactile sensing, Cai et al. developed a neuromorphic framework that integrates IGZO-based first-spike-timing (FST) tactile encoders with coupling-enhanced transistor synapses [Figure 5B(i)]. This architecture can efficiently convert continuous pressure stimuli into precise FST spike patterns, thereby providing an energy-efficient sensory front-end for downstream spiking neural networks (SNNs). When applied to spatiotemporal tactile-feedback tasks, including autonomous vehicle status detection and intelligent robotic navigation [Figure 5B(ii)], the system achieved high recognition accuracies of 98.4% and 98.2%, respectively, while reducing network training time by 90.9%^[136]. In the visual domain, Song et al. proposed an artificial optoelectronic synapse utilizing a source-sharing circuitry design. Under dynamic irradiation, the device intrinsically integrates and decodes temporal light patterns, proving that a-IGZO circuitry can perform complex feature extraction directly at the sensory terminal before central routing^[137].

Furthermore, near-sensor computing proves exceptionally effective in complex environmental chemical monitoring. Addressing the poor selectivity of traditional gas sensors, Chu et al. constructed a power-efficient gas-sensing synaptic diode based on a lateral pentacene/a-IGZO p-n junction^[138]. When exposed to the target environment, different gas molecules (such as NO₂ and NH₃) act as distinct “chemical dopants” adsorbing at the organic/inorganic interface. This dynamic adsorption not only modulates the channel conductance but also regulates the update rate of synaptic weights (i.e., the learning rate). Inspired by the signal-processing behavior of the biological olfactory system, the device integrates gas sensing and synaptic processing within a single heterojunction platform, enabling the direct discrimination of mixed gases at the sensing node and providing a promising hardware solution for highly selective flexible electronic noses^[138]. Near-sensor a-IGZO circuits can therefore perform local amplification, filtering, spike encoding, and sensory feature extraction. Their main advantages are reduced data transmission and compatibility with large-area flexible electronics, whereas circuit gain, device uniformity, mechanical drift, and peripheral power remain important system-level constraints.

RC

RC is particularly suited to temporal inputs that require nonlinear transformation and short-term memory, such as physiological waveforms, motion sequences, and speech. Unlike recurrent neural networks trained through backpropagation through time, RC uses a fixed nonlinear reservoir and trains only the readout layer. This reduces training complexity, although the performance still depends on the reservoir dynamics, input encoding, and stability of the readout states^[139].

At the physical implementation level, the migration behavior of electrolyte ions or the charge trapping/release process of interface defects in a-IGZO synaptic transistors endows the devices with natural “fading memory” characteristics, which are core elements for building a physical reservoir. Yang et al. used the ionic relaxation of an EDL-coupled InGaZnO artificial synapse as a time-multiplexed physical reservoir. The device provided nonlinear fading-memory states, while the spoken-digit classification was performed using a trained readout layer. Under the reported conditions, the system achieved 100% accuracy, with an average device power of approximately 9.3 nW and an estimated energy of approximately 1.9 nJ per reservoir state^[140].

Furthermore, in RC, precisely manipulating the retention time to achieve fading memory is crucial for matching the timescale of dynamic sequential inputs. To meet this requirement, Jang et al. developed a physical reservoir based on leaky two-transistor (2T) DRAM nodes utilizing nanometer-thick a-IGZO films [Figure 5C(i)]. By exploiting the controllable retention loss and PPF behavior of the 2T DRAM, the system efficiently mapped time-varying inputs into high-dimensional reservoir states through a 4-bit pulse scheme. When applied to handwritten digit recognition [Figure 5C(ii)], the framework achieved a high recognition accuracy of 95.06% within only 20 training epochs, demonstrating its potential for fast and energy-efficient temporal information processing^[141].

To cope with more complex, higher-bandwidth scenarios, RC systems are evolving toward optoelectronic multimodal directions, using the recombination dynamics of photo-generated carriers instead of slow ion migration. Cui et al. designed a bioinspired in-materia analog photoelectronic RC system. By utilizing a receptive field-inspired encoding scheme and the intrinsic dynamics of photogenerated carriers, this system achieved over 90% recognition accuracy across four distinct human motion datasets. Remarkably, it verified falling behavior recognition with an exceptionally low energy consumption of ~45.78 μJ per action, significantly outperforming traditional vision processing hardware^[142]. Fu et al. synthesized IGZO/PVP composite nanofiber neuromorphic transistors. This unique nanofiber network constructed a highly efficient reservoir for spatiotemporal visual tasks, achieving 92.86% accuracy in hand gesture recognition with a microscopic energy footprint of just 49 pJ per reservoir state^[88].

In flexible implementations, changes in device response caused by mechanical deformation may alter the reservoir states. Zhang et al. showed that the recognition performance of a flexible RC system could be recovered after bending by retraining only the linear readout layer^[33]. This approach does not eliminate mechanically induced drift, but it provides a relatively simple means of compensating for reproducible changes in reservoir response. Overall, a-IGZO devices offer fading memory and nonlinear dynamics suitable for temporal processing, although task performance remains sensitive to relaxation time, environmental drift, input encoding, and readout-layer training.

CIM

CIM reduces repeated data transfer between separate memory and processing units by performing multiply–accumulate operations within or close to the memory array^[143]. In analog crossbars, multiplication is represented by conductance-dependent current and summation by current accumulation along shared lines. However, practical CIM systems still require programming, control, sensing, and data-conversion circuits, and therefore do not eliminate all data-movement or peripheral-circuit overhead. 1T1R arrays for resistive memories, 2T1C and 3T1C eDRAM cells, and modified 6T SRAM cells for computing-in-memory applications^[143]. In these structures, a TFT provides cell selection, programming-current control, or charge isolation, depending on the memory architecture.

Compared with two-terminal memristors, three-terminal a-IGZO synaptic transistors separate gate-controlled programming from source–drain readout, allowing more controllable multilevel updates and reduced read disturbance. Their larger cell area and wiring overhead reduce integration density, whereas memristors offer compact, nonvolatile cross-point storage but often suffer from nonlinear or asymmetric updates, device and cycle variability, and sneak-path currents^[144]. Thus, a-IGZO transistors favor precise weight control and active-matrix addressing, while memristors favor high-density integration and nonvolatile retention.

In passive resistive-memory crossbars, sneak-path currents reduce the read margin and can restrict practical array scaling. Active 1T1R arrays address this problem by using a transistor to isolate unselected cells^[145,146]. In such configurations, the a-IGZO TFT primarily serves as an access device rather than the analog weight-storage element. This improves array selectivity but does not directly use the transistor’s own multilevel synaptic characteristics.

To further avoid the inherent non-idealities of memristors (such as nonlinear writing and asymmetry), the capacitive in-memory computing (C-CIM) paradigm, based on charge retention in capacitors, has received extensive attention. In this architecture, synaptic weights are stored on capacitors as analog voltages rather than relying on resistive switching of materials. Pioneering this domain, Won et al. performed a device-algorithm co-optimization for an on-chip trainable capacitor-based synaptic device utilizing IGZO TFTs, ensuring highly robust weight retention [Figure 5D(i) and (ii)]^[147]. Building upon this approach, Kang et al. designed an advanced analog neuromorphic on-chip training system based on an IGZO TFT 6T1C synaptic memory structure. Exploiting the extremely low leakage current of IGZO, they achieved a 10,000-fold enhanced retention time. This 6T1C architecture delivered an average of 367 distinct analog states with an extremely high R-squared linearity of 0.99, enabling the system to conduct successful analog on-chip training with an accuracy of 97.1% on the MNIST dataset^[148]. Furthermore, seeking to overcome the density and leakage bottlenecks of traditional CMOS embedded DRAM, Kim et al. proposed a highly efficient compute-in-memory array design using 3T-based stacked hybrid IGZO/Si eDRAM cells. This vertical integration harnesses the exceptionally long retention time of IGZO (> 100 s) to drastically minimize refresh cycles while significantly increasing memory cell density for demanding computational workloads^[149].

However, deploying CIM architectures on flexible electronics faces severe reliability challenges due to inherent device variations. Systematic simulation studies by Park et al. showed that spatial variations and threshold voltage drift of IGZO TFTs within the array critically compromise the consistency of analog VMM operations. Since analog computing relies heavily on uniform device I-V characteristics, this D2D variation directly translates into systematic errors. Their simulations quantitatively demonstrated that extensive threshold voltage fluctuations profoundly degrade the inference accuracy of SNNs^[150]. To address these limitations, the same group subsequently developed high-performance IGZO synapse arrays using an ITO floating gate. This architecture significantly suppressed charge leakage and minimized D2D variations, ultimately achieving a 98.31% MNIST inference accuracy in an 8 × 8 array SNN simulation^[28]. Beyond stabilizing synaptic weights, robust spiking neurons are essential for complete flexible SNN hardware. To this end, Lebanov et al. proposed fully flexible IF neuron circuits based exclusively on unipolar IGZO TFTs. Capitalizing on a-IGZO’s ultra-low leakage, these circuits effectively suppressed current drift and maintained highly stable membrane potential integration, establishing a crucial hardware foundation for highly reliable flexible SNNs^[134].

a-IGZO TFTs are therefore attractive for active-matrix addressing, low-leakage charge storage, and controllable analog weight updates in CIM circuits. Their larger cell area relative to two-terminal memories, together with array nonuniformity, peripheral-circuit overhead, and strain-induced parameter drift, remains a limitation for high-density flexible integration. Further progress requires coordinated optimization of the memory cell, access transistor, interconnects, and training algorithm.