Traditional AI hardware suffers from high energy consumption and latency due to the separation of sensing, processing, and memory. Biological nervous systems are significantly more energy-efficient due to the co-location of these functionalities. This paper aims to emulate this biological efficiency by developing a single device capable of performing all three functions. Current approaches, such as those using 2D materials or phase-change materials, lack either sensing or non-volatile memory capabilities. Organic electrochemical transistors (OECTs) offer a potential solution due to their low power consumption and ability to operate in wet environments, but previous iterations have been limited to either sensing or memory functions, not both simultaneously. This research seeks to overcome this limitation by creating a dual-mode OECT.

Existing literature highlights the energy efficiency of biological nervous systems compared to silicon-based AI hardware, emphasizing the need for biologically inspired hardware. Previous attempts to create artificial systems that mimic biological sensing, processing and memory have relied on heterogeneous module integration, leading to challenges in fabrication, integration density, and conductance matching. While some progress has been made using 2D materials and phase-change or redox memristors, these devices fall short in terms of providing a unified platform for sensing, non-volatile memory, and processing. OECTs show promise due to their low power consumption and ability to operate in wet environments. However, existing OECT designs typically lack either multi-modal sensing capabilities or non-volatile memory functions. The challenge lies in achieving both volatile and non-volatile behavior within a single OECT for seamless integration.

The researchers designed a vertical traverse OECT (v-OECT) with a high depth/length ratio to minimize electric field gradients and prevent ion drift. The channel is composed of a crystalline-amorphous PTBT-p material, enabling selective ion doping. Crystallinity was controlled through thermal annealing, influencing the ion trapping and mobility. A polarizable gold gate electrode was used to prevent counterion compensation and maintain non-volatility. The volatile operation mode was achieved by applying a low-gate potential (LGP), doping the amorphous region and allowing rapid ion diffusion. The non-volatile mode was achieved using a high-gate potential (HGP), trapping ions in the crystalline region. The device was characterized using various techniques including transfer curves, UV-vis spectroscopy, X-ray scattering, and X-ray photoelectron spectroscopy (XPS). Multi-modal sensing capabilities were demonstrated by measuring plant ion concentration changes, ECG signals, temperature, and taste stimuli. Non-volatile synaptic behavior was evaluated by assessing state retention, switching stochasticity, and the number of distinct conductance states. Spike-timing-dependent plasticity (STDP) was demonstrated using a homogeneous 1T1R architecture composed of two cv-OECTs. Finally, reservoir computing simulations were used to assess the potential for real-time cardiac disease diagnosis using an array of these OECTs.

The researchers successfully fabricated a dual-mode OECT capable of both volatile receptor and non-volatile synaptic functions. The v-OECT exhibited high sensitivity and fast switching speeds in its volatile receptor mode, demonstrating successful multi-modal sensing applications such as detecting ion concentration changes in plants, recording ECG signals, and sensing temperature and taste. In the non-volatile synaptic mode, the device showed 1024 (10-bit) distinct analogue states, a wide dynamic range, and excellent state retention exceeding 10,000 s. The homogeneous integration of multiple cv-OECTs demonstrated the successful implementation of STDP learning rules, mimicking fundamental learning processes in the brain. Moreover, simulations demonstrated the potential of the integrated cv-OECTs for real-time cardiac disease diagnosis using reservoir computing, reaching 100% accuracy after 700 training epochs. The on/off ratio of the cv-OECT was as high as 8 × 10⁶, which is a record-high value in OMIEC-based transistors. The switching times were shorter than that of p-OECT with a similar footprint (>100 ms) because of the smaller doping area of v-OECT. The subthreshold swing (SS) value was close to the thermodynamic limit (59.6 mV dec⁻¹) in the subthreshold regime with VDS = 0 V.

The results of this study demonstrate the successful creation of a versatile organic electrochemical transistor that performs both multi-modal sensing and non-volatile memory functions. This advance has significant implications for the development of more energy-efficient and biologically inspired AI systems. The high sensitivity, fast response, and high on/off ratio of the volatile receptor mode, combined with the high-fidelity, long-term stability, and high-bit-resolution memory capability of the non-volatile synaptic mode, make this transistor highly suitable for a variety of sensing and in-memory computing applications. The demonstration of STDP functionality and simulated performance in real-time cardiac diagnosis further validates the potential of homogeneous integration of this device for advanced applications. The findings bridge the gap between biological systems and artificial intelligence, paving the way for more sophisticated and efficient neuromorphic computing.

This research successfully demonstrated a novel organic electrochemical transistor with reconfigurable multi-modal sensing and non-volatile analogue memory capabilities. By controlling the device architecture, channel microstructure, and electrode process, the researchers achieved a single device capable of both volatile and non-volatile operations. Homogeneous integration of these transistors enabled the implementation of advanced functionalities such as conditioned reflexes and real-time cardiac disease diagnosis. Future research could focus on further scaling down the device dimensions to improve speed and energy efficiency, as well as exploring different channel materials and architectures to broaden the range of applications.

The study primarily focused on simulations for the complex applications such as cardiac disease diagnosis. Further experimental validation of these complex applications is needed. While the device demonstrated excellent performance, the long-term stability and reliability under diverse environmental conditions warrant further investigation. The effects of device-to-device variations within arrays on the overall system performance should also be investigated in more detail. Finally, the scalability of the fabrication process to large-scale production needs to be addressed for practical deployment.