Single Transistor Unifies Reservoir and Readout for Machine Vision

Researchers have demonstrated a vertical organic field-effect transistor that monolithically integrates a volatile, nonlinear optical reservoir with a linear, nonvolatile electrical readout within a single device footprint. Published in Light: Science & Applications, the study addresses a fundamental hardware conflict in analog reservoir computing, where the reservoir requires volatile, nonlinear dynamics to process temporal information, while the readout layer demands linear, symmetric, and nonvolatile conductance modulation for stable weight storage. By physically decoupling photocarrier dynamics from ferroelectric polarization, the team created a dual-mode architecture that allows for the orthogonal optimization of these conflicting requirements without the performance trade-offs typical of single-device multifunctional systems.

The device, termed a ferroelectric vertical organic field-effect transistor or Fe-vOFET, utilizes a vertical-channel geometry to amplify optical feedback and achieve a high conductance on/off ratio of approximately 10^7. This geometric confinement is critical for high-fidelity weight mapping, as it provides the strong field necessary to distinguish between different conductance states. The architecture supports two complementary operational modes: an optical reservoir mode governed by the photocarrier dynamics of a Rubrene/F16CuPc heterojunction, and an electrical-readout mode programmed via ferroelectric switching in a metal–ferroelectric–metal–insulator–semiconductor stack. This structural separation ensures that the reservoir dynamics and memory retention can be optimized independently within the same physical structure.

In the optical mode, the device leverages a broadband-absorbing heterojunction to generate wavelength-dependent relaxation transients. When subjected to optical pulse trains across the ultraviolet to near-infrared spectrum, the Fe-vOFET exhibits nonlinear conductance evolution with volatile dynamics. The relaxation kinetics are intrinsically governed by the incident photon energy, meaning that stimuli at different wavelengths trigger distinct decay profiles. This spectral diversity allows a single device to project multi-wavelength optical sequences into orthogonal reservoir subspaces, effectively enabling multi-wavelength grouped reservoir computing. The vertical channel, with a length of approximately 255 nm set by the active-layer thickness, accelerates photocarrier separation and suppresses non-radiative recombination, thereby enhancing the feedback strength and broadening the spatial response profile.

To quantify the short-term memory behavior, the researchers performed paired-pulse facilitation and excitatory postsynaptic current measurements at 340, 530, 625, and 780 nm. Double-exponential fitting of the decay curves yielded characteristic time constants ranging from 0.182 to 0.347 seconds for the fast component and 1.51 to 3.61 seconds for the slow component, depending on the wavelength. These results confirm that the device possesses multi-timescale fading-memory dynamics suitable for temporal information processing. The system successfully resolved 64 distinct conductance levels in response to six-bit optical pulse trains, demonstrating its ability to map complex optical inputs into a broad spectrum of distinguishable transient memory states, although some neighboring states remained relatively close, suggesting room for further interface engineering.

The electrical readout mode relies on a robust MFMIS gate stack designed to mitigate interface trapping, a common issue in conventional ferroelectric/semiconductor junctions that leads to incomplete polarization switching and retention degradation. By introducing an intermediate floating gate and a buffering dielectric, the MFMIS configuration decouples the ferroelectric layer from the organic channel. This design enhances electrostatic coupling and allows for the faithful transduction of ferroelectric polarization switching into highly linear and symmetric conductance updates. The device exhibits a clear counterclockwise hysteresis with a memory window of approximately 8 V, which is highly symmetric around 0 V, indicating balanced polarization switching with negligible contribution from interfacial charge trapping.

Reliability measurements underscore the robustness of the nonvolatile readout. Retention tests demonstrate that the ON and OFF states remain distinct and stable for over 3000 seconds, validating the device’s capability for long-term synaptic weight storage. In endurance assessments, the device sustains a robust memory window of greater than 10^5 after 10^6 program/erase cycles at ±5 V and 200 μs pulse widths. Quantitative analysis of the analog conductance modulation reveals an asymmetry nonlinearity of 0.31 ± 0.05, with potentiation and depression parameters of 0.80 ± 0.04 and 1.15 ± 0.08, respectively. These attributes confirm that the MFMIS stack successfully transduces polarization dynamics into precise weight updates, eliminating the need for complex iterative write-verify circuitry often required in other neuromorphic devices.

To validate the practical relevance of the Fe-vOFET, the team implemented a traffic-trajectory recognition task using a custom video dataset comprising eight motion categories. The Fe-vOFET-based sensor array served as the front-end encoder, generating temporally compressed states that integrate optical responses over the full motion sequence. Unlike conventional image sensors that record only instantaneous spatial information, the Fe-vOFET frames preserve the motion contour in a single compressed frame, enabling simultaneous encoding of spatial structure and temporal evolution. This spatiotemporal fusion led to a final recognition accuracy of 97.2%, whereas a baseline configuration using conventional image frames saturated at only approximately 40% under the same network architecture and training protocol.

The study further benchmarked the device on the Weizmann human-action dataset, a standard task for assessing spatiotemporal pattern recognition. Video sequences were encoded as optical pulse trains and distributed to the Fe-vOFET via four parallel spectral paths at 340, 530, 625, and 780 nm, creating four distinct sub-reservoirs. The output of each virtual node was derived directly from experimentally measured optical conductance responses to ensure physical fidelity. Principal component analysis was performed to extract 40 principal components from each sub-reservoir, which were then concatenated into a 160-dimensional vector for a single-layer artificial neural network readout. The readout weights were optimized using the Adam algorithm and mapped onto the experimentally extracted long-term potentiation and long-term depression characteristics of the Fe-vOFET.

Classification results highlight the critical role of multi-timescale reservoir dynamics. A reservoir-free baseline, lacking temporal accumulation, achieved a recognition rate of only 67.8%. Restricting the system to a single sub-reservoir, equivalent to a single-timescale condition, resulted in an accuracy of 85.46%. As additional sub-reservoir groups were introduced, the accuracy increased, reaching 91.95% in the four-group configuration. This comparison indicates that the multi-timescale dynamics provide complementary temporal features that are beneficial for improving computational performance. The accuracy achieved using the physical LTP/LTD characteristics of the device was comparable to that of software-optimized weights, confirming that the high linearity and symmetry enabled by the MFMIS readout are sufficient to support precise learning.

The framework was also validated on static image benchmarks, including MNIST handwritten-digit recognition and CIFAR-10 color-image recognition. For CIFAR-10, HOG features were extracted from RGB channels and encoded into reservoir inputs using a 4-bit mapping scheme. Despite the increased complexity of color images, the system maintained good recognition performance on four representative categories. An order-of-magnitude estimate of the readout-side electrical energy suggests approximately 10 nJ per inference for the output layer, corresponding to about 10 pJ per cell operation, though this estimate excludes peripheral circuitry and optical source losses. The work establishes a scalable device paradigm that unifies dynamical richness with programming precision, paving the way for energy-efficient, all-in-one dynamic machine vision systems.

The device fabrication process involved a sequential multistep approach on silicon substrates, beginning with the deposition of a 20 nm titanium nitride bottom gate followed by a 16 nm hafnia-zirconia layer via atomic layer deposition. A 20 nm titanium nitride floating gate was then added, and the structure underwent rapid thermal annealing at 450 °C for 30 seconds under nitrogen to crystallize the ferroelectric layer. A 10 nm hafnia interfacial dielectric completed the metal–ferroelectric–metal–insulator–semiconductor heterostructure. Monolayer graphene was mechanically exfoliated and transferred onto the surface to serve as the shared bottom source electrode, while high-quality rubrene single crystals were synthesized via physical vapor transport and transferred onto the graphene stack. A 10 nm copper hexadecafluorophthalocyanine film was thermally evaporated onto the rubrene to form the organic heterojunction, and top electrodes were laminated to finalize the transistor architecture.

Electrical and photoelectrical characterizations were conducted using a semiconductor parameter analyzer within a closed-loop low-temperature probe station under a vacuum pressure of approximately 10−2 Pa. Optical pulse stimulation was provided by fiber-coupled light-emitting diodes at specific wavelengths of 340, 530, 625, and 780 nm, with irradiance calibrated to ensure consistent optical power density across different illumination conditions. Surface morphologies of the organic crystals and thin films were analyzed using atomic force microscopy in tapping mode. Ferroelectric properties were quantitatively assessed using piezo response force microscopy and a dedicated ferroelectric test system, which recorded polarization–electric field hysteresis loops to evaluate remnant polarization, coercive field, and switching stability. These combined measurements provided comprehensive insight into the coupling mechanisms between the optical and electrical modes of the device.

To ensure the scalability of the architecture, the team evaluated device-to-device variability across 21 independently fabricated transistors. The statistical distribution revealed highly stable memory windows with a median on/off ratio of approximately 10^6 and a variation coefficient below 15%. This observed variation was primarily attributed to minor thickness fluctuations in the rubrene single crystals and the transferred graphene layers. The researchers noted that standardizing the transfer process and crystal selection could further suppress this variability for large-scale integration. Additionally, the device maintained clear light-triggered transient responses even at an ultralow drain bias of −25 μV, highlighting its potential for ultralow-power optical reservoir operation. The effective operating timescale of the optical reservoir was in the tens-of-milliseconds regime, determined by the interplay of photocarrier trapping and relaxation kinetics, which govern the fading-memory behavior required for reservoir computing.

Subject of Research: Orthogonally regulated multi-timescale reservoir and linear nonvolatile readout in a vertical transistor for dynamic machine vision

Article Title: Orthogonally regulated multi-timescale reservoir and linear nonvolatile readout in a vertical transistor for dynamic machine vision

Article References: Dai, Q., Cui, H., Cao, Y., Chen, K., Ai, C., Wang, Q., Zhu, S., Wang, W., Fang, Z., Yao, B., Pei, M., Shi, Y., & Li, Y. (2026). Orthogonally regulated multi-timescale reservoir and linear nonvolatile readout in a vertical transistor for dynamic machine vision. Light: Science & Applications, 15(1), Article 398. https://doi.org/10.1038/s41377-026-02420-4

Image Credits: AI Generated

DOI: 10.1038/s41377-026-02420-4

Keywords: Orthogonally, regulated, multi-timescale, reservoir, linear, nonvolatile, readout, vertical, transistor, dynamic, machine, vision