The von Neumann architecture bottleneck in artificial intelligence has spurred the development of hardware artificial neural networks (HW-ANNs). While recurrent neural networks (RNNs) excel at processing spatiotemporal signals, their training complexities limit their applicability. Reservoir computing (RC) offers a solution by requiring training only the output layer weights, significantly reducing training costs. However, existing neuromorphic hardware mainly focuses on shallow RC with limited spatial and temporal scales, hindering performance in complex spatiotemporal tasks. This research addresses these limitations by introducing a novel organic neuromorphic vertical transistor with distributed reservoir states, inspired by the distributed memory mechanisms in primate brains. This device aims to enrich the reservoir state space and improve the range ratio of spatial and temporal characteristics for more effective signal processing.

Numerous neuromorphic devices have been applied to reservoir computing, but most focus on shallow RC with monotonic reservoir state spaces. This limitation stems from the reliance on monotonic carrier dynamics, leading to narrow range ratios of spatial and temporal characteristics. This restricts the richness of the reservoir state space, causing overlap and reducing recognition accuracy and prediction correlation. While increasing input signal modes can improve temporal characteristics in shallow-RC, the spatial limitations persist, increasing signal error and preprocessing costs. The primate brain's distributed memory processing, where different brain regions handle different information aspects before integration, serves as a model for a more efficient and accurate system. This distributed memory concept inspires the design of a physical device with distributed reservoir states to facilitate the distributed mapping of spatiotemporal signals.

This study introduces an organic neuromorphic vertical field-effect transistor with distributed reservoir states (VOFET-DR) as the reservoir node. The device employs a p-n organic semiconductor bulk heterojunction (BHJ) active layer (N2200:POFDIID) and an ultra-short channel vertical architecture. The vertical architecture significantly enhances the feedback intensity by reducing the distance for carrier transport. Three operating modes are used: inputting voltage sequence pulse signals, inputting laser sequence pulse signals, and simultaneously applying gate bias while inputting laser sequence pulse signals. The device's characteristics, including transfer curves, hysteresis windows, short-term memory currents (in both voltage and light pulse modes), and nonlinear temporal characteristics (τ), are investigated. Feedback strength (FE and Fph) is also analyzed. A 6-bit light sequence signal test and a 64-binary timing signal test evaluate the device's ability to map multi-bit signals. The effect of gate bias (VGS) on the device's performance is explored to demonstrate the distributed reservoir states, analyzing nonlinear temporal characteristics, feedback strength, and the impact on mapping sequence signals. The underlying mechanism is explained through analysis of charge trapping, activation energy, surface potential distribution, Schottky barrier modulation, and charge distribution using COMSOL simulations. Two applications are presented: satellite remote sensing image recognition and traffic trajectory prediction. For image recognition, the device processes UV, Vis, and NIR spectral features independently, with each feature fed to parallel sub-reservoirs with different VGS biases. The outputs are then integrated and trained using a fully connected network. For traffic trajectory prediction, the device combined with optical flow and inter-frame difference methods is used to extract motion information from traffic scenes. The system uses the velocity and coordinate information of the object in the previous three frames as input to predict the next frame's position.

The VOFET-DR device demonstrates a distributed reservoir state space with 1152 states, significantly exceeding the capacity of shallow-RC systems. The device exhibits ultra-wide range ratios of temporal (2640) and spatial (650) characteristics, surpassing reported neuromorphic devices. The grouped-RC network, based on the VOFET-DR, achieves over 94% recognition accuracy in satellite remote sensing image recognition and over 95% prediction correlation in traffic trajectory prediction. These high accuracies are achieved with a significantly reduced number of weights (over 90% fewer) compared to traditional ANN and CNN architectures. The analysis of the device mechanism reveals that the bulk heterojunction structure, ultra-short channel, and gate bias modulation contribute to the rich reservoir states. The energy barrier for charge trapping is wavelength-dependent, enabling the device to capture different physical characteristics of input signals. The gate bias efficiently adjusts charge injection and the electric field distribution, further enriching the carrier dynamics. COMSOL simulations support these findings. The negligible additional power density from gate control underscores the device's energy efficiency.

The findings demonstrate the success of using a novel organic neuromorphic transistor with distributed reservoir states to overcome the limitations of shallow RC. The high accuracy and efficiency in both image recognition and trajectory prediction showcase the potential of grouped-RC for complex spatiotemporal tasks. The significant reduction in the number of weights needed compared to conventional ANNs and CNNs highlights the computational efficiency of the approach. These results validate the effectiveness of the design inspired by the primate brain's distributed memory mechanisms. The broad applicability of the device in different domains further suggests its potential as a fundamental building block for future neuromorphic hardware.

This research presents a groundbreaking organic neuromorphic vertical transistor with distributed reservoir states for grouped-RC. The device successfully addresses the limitations of shallow RC by achieving a significantly richer reservoir state space, ultra-wide range ratios of temporal and spatial characteristics, and negligible additional power density. The high accuracy in image recognition and trajectory prediction tasks, along with a drastically reduced number of weights, demonstrates the potential of this approach for developing high-performance and energy-efficient neuromorphic computing systems. Future work could focus on scaling up the device for larger-scale applications and exploring further optimizations for even higher accuracy and efficiency.

While the presented results demonstrate significant improvements, several limitations exist. The current study is limited to specific applications (satellite image recognition and traffic trajectory prediction). Further research should investigate the device's performance across a broader range of tasks and datasets. The device fabrication process currently involves multiple steps, potentially impacting scalability. Future research should explore simplified fabrication methods to enable higher-throughput production. Finally, the study's focus on specific types of organic semiconductors might warrant investigation into the performance of other materials.