image: Figure 1 | Reconfigurable photoelectric dual-mode Fe-vOFET for neuromorphic computing. a, Schematic of an Fe-vOFET–based photoelectric RC system for human action recognition: a front-end reservoir for nonlinear temporal mapping and a back-end ANN readout for classification, both realized with the same device in distinct modes. b, Optically driven operation: broadband excitation yields wavelength-dependent relaxation, enriching reservoir states for multi-wavelength grouped RC. c Electrically driven operation: reversible ferroelectric polarization enables linear, symmetric, nonvolatile conductance updates suitable for precise weight programming and stable inference.
Credit: Yun Li et al.
Artificial intelligence is rapidly transforming machine vision, robotics, autonomous driving, and intelligent surveillance. However, recognizing dynamic visual information, such as human motion or moving traffic trajectories, remains computationally demanding. Conventional deep neural networks usually require large numbers of trainable parameters and high energy consumption, making them difficult to deploy in compact edge devices.
Reservoir computing provides an attractive alternative for temporal information processing. In this computing framework, a reservoir transforms time-dependent inputs into high-dimensional dynamic states, while a simple readout layer performs classification. This greatly reduces training complexity. Yet, building reservoir computing directly in hardware faces a major challenge: the reservoir and the readout require very different physical properties. The reservoir should be nonlinear and volatile so that it can capture fading memory, whereas the readout should be linear and nonvolatile so that it can store weights precisely and stably.
In a new paper, a team of scientists led by Professor Yun Li and Professor Yi Shi from Nanjing University, Dr. Mengjiao Pei from The Hong Kong Polytechnic University, and Professor Bo Yao from Shaoxing University, and co-workers have developed a ferroelectric vertical organic field-effect transistor that addresses this conflict within a single device platform. The device, termed a Fe-vOFET, combines a vertical organic transistor structure with a ferroelectric gate stack. It operates in two distinct modes: an optically driven reservoir mode and an electrically driven readout mode.
In the optical mode, a Rubrene/F16CuPc organic heterojunction responds to light pulses with wavelength-dependent relaxation dynamics. Different wavelengths generate different fading-memory timescales, allowing the same device platform to form multi-timescale sub-reservoirs. These sub-reservoirs enrich the temporal feature space and improve the separation of dynamic input sequences. In the electrical mode, a metal-ferroelectric-metal-insulator-semiconductor gate stack converts ferroelectric polarization into linear, symmetric, and nonvolatile conductance updates, providing the precise readout required for stable learning. The scientists summarize the key design concept of their device:
“We use photocarrier dynamics to build a volatile optical reservoir, while ferroelectric polarization is used to independently program a nonvolatile electrical readout,” the scientists explained. “In this way, the two key components of reservoir computing, which normally require conflicting device properties, can be optimized separately within one vertical transistor platform.”
They added: “The vertical geometry is critical to this design. The short vertical channel enhances optical feedback and strengthens the transient photo response, while also providing a large conductance window with an on/off ratio of about ten million. This is important for accurate weight mapping in the readout layer.” The device also exhibits reliable nonvolatile memory characteristics, including retention for more than 3000 s and endurance over one million switching cycles.
To demonstrate the practical value of this platform, the researchers applied the Fe-vOFET to dynamic machine-vision tasks. “In a traffic-trajectory recognition task, the bio-inspired vision sensor can compress motion information into a single spatiotemporal frame and achieve 97.2% recognition accuracy, greatly outperforming a conventional frame-based sensing approach,” they said. In a standard human-action recognition benchmark based on the Weizmann dataset, the multi-wavelength grouped reservoir computing system achieved over 90% accuracy, confirming the advantage of multi-timescale reservoir dynamics for motion information processing.
“By resolving the conflict between volatile reservoir dynamics and nonvolatile readout, this vertical transistor architecture provides a device-level route to combine rich temporal dynamics with precise weight storage,” the scientists forecast. “This strategy may help advance energy-efficient, highly integrated, and adaptive dynamic machine-vision hardware for future edge-intelligence applications.”
Journal
Light: Science & Applications
Article Title
Orthogonally Regulated Multi-Timescale Reservoir and Linear Nonvolatile Readout in a Vertical Transistor for Dynamic Machine Vision