Orthogonally regulated multi-timescale reservoir and linear nonvolatile readout in a vertical transistor for dynamic machine vision
Abstract
Implementing analog reservoir computing (RC) at the edge is fundamentally constrained by conflicting hardware requirements for the reservoir (volatile, nonlinear) and readout (nonvolatile, linear) layers. Here, we resolve this dilemma by monolithically integrating a multi-timescale photoelectric reservoir and a linear ferroelectric readout within a single vertical organic field-effect transistor (Fe-vOFET). This architecture functionally decouples photocarrier dynamics from ferroelectric polarization, allowing orthogonal optimization of reservoir dynamics and synaptic weights. In the optical mode, the Rubrene/F ₁₆ CuPc heterojunction underpins reservoir dynamics, exploiting wavelength-dependent relaxation kinetics to construct multi-timescale grouped reservoirs that project inputs into high-dimensional spatiotemporal states. Conversely, in the electrical mode, the metal–ferroelectric–metal–insulator–semiconductor (MFMIS) stack governs readout functionality by mitigating interface trapping, translating remnant polarization into linear, symmetric, nonvolatile weight updates (endurance >10 ⁶ , retention >3000 s, asymmetry nonlinearity 0.31 ± 0.05). The vertical-channel geometry is pivotal, simultaneously amplifying optical feedback and boosting the on/off ratio to ~10 ⁷ , which is essential for high-fidelity weight mapping. Leveraging this dual-mode operation, we demonstrate a compact RC system achieving >90% accuracy in dynamic human-action recognition. This 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.
Related
- Nonabelian vortices: From topological strings to nonvolatile memory
- Spatiotemporal programming via asymmetric dielectric engineering for nonvolatile 2D optoelectronics
- The Ericson Transition in Time-Reversal Invariant Systems: Orthogonally Invariant Hamiltonians
- The Global Topology of Orthogonally Decomposable Tensor Landscapes
- Dynamical mean-field limit and replica-symmetric free energy for the orthogonally-invariant SK model
BibTeX
@article{dai2026orthogonally,
title = {{Orthogonally regulated multi-timescale reservoir and linear nonvolatile readout in a vertical transistor for dynamic machine vision}},
author = {Qinyong Dai and Hangyuan Cui and Yishan Cao and Kai Chen and Chao Ai and Qijing Wang and Shunwei Zhu and Wenhao Wang and Zebo Fang and Bo Yao and Mengjiao Pei and Yi Shi and Yun Li},
journal = {Light: Science \& Applications},
year = {2026},
month = nov,
doi = {10.1038/s41377-026-02420-4},
url = {https://www.nature.com/articles/s41377-026-02420-4},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Orthogonally regulated multi-timescale reservoir and linear nonvolatile readout in a vertical transistor for dynamic machine vision
AU - Qinyong Dai
AU - Hangyuan Cui
AU - Yishan Cao
AU - Kai Chen
AU - Chao Ai
AU - Qijing Wang
AU - Shunwei Zhu
AU - Wenhao Wang
AU - Zebo Fang
AU - Bo Yao
AU - Mengjiao Pei
AU - Yi Shi
AU - Yun Li
PY - 2026
DA - 2026/11/09
JO - Light: Science & Applications
DO - 10.1038/s41377-026-02420-4
UR - https://www.nature.com/articles/s41377-026-02420-4
AB - Implementing analog reservoir computing (RC) at the edge is fundamentally constrained by conflicting hardware requirements for the reservoir (volatile, nonlinear) and readout (nonvolatile, linear) layers. Here, we resolve this dilemma by monolithically integrating a multi-timescale photoelectric reservoir and a linear ferroelectric readout within a single vertical organic field-effect transistor (Fe-vOFET). This architecture functionally decouples photocarrier dynamics from ferroelectric polarization, allowing orthogonal optimization of reservoir dynamics and synaptic weights. In the optical mode, the Rubrene/F ₁₆ CuPc heterojunction underpins reservoir dynamics, exploiting wavelength-dependent relaxation kinetics to construct multi-timescale grouped reservoirs that project inputs into high-dimensional spatiotemporal states. Conversely, in the electrical mode, the metal–ferroelectric–metal–insulator–semiconductor (MFMIS) stack governs readout functionality by mitigating interface trapping, translating remnant polarization into linear, symmetric, nonvolatile weight updates (endurance >10 ⁶ , retention >3000 s, asymmetry nonlinearity 0.31 ± 0.05). The vertical-channel geometry is pivotal, simultaneously amplifying optical feedback and boosting the on/off ratio to ~10 ⁷ , which is essential for high-fidelity weight mapping. Leveraging this dual-mode operation, we demonstrate a compact RC system achieving >90% accuracy in dynamic human-action recognition. This 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.
SN - 2047-7538
ER -