In recent years, artificial intelligence (AI) based on compute-in-memory (CIM) integrated artificial neural networks has experienced rapid development. Large-scale neural network arrays and computational models have yielded significantly enhanced computing power. However, current artificial neural networks still lag far behind biological nervous systems in terms of energy efficiency and functionality. The latter’s multimodal sensory-memory-computing integrated architecture enables it to function as an intelligent system capable of real-time integrated information perception, processing and feedback, endowed with sophisticated intelligent functions such as conditioned reflexes. Therefore, the development of biomimetic learning hardware represents an urgent imperative for advancing AI systems, with the core challenge lying in hardware-level emulation of peripheral neural receptors as well as central neural synapses and neurons.
Nevertheless, the heterogeneous material and device architectures inherent to conventional multi-module integration strategies impose constraints on processing compatibility, conductance matching and integration density, rendering them unsuitable for the construction of miniaturized AI systems. Thus, the implementation of multimodal sensing and compute-in-memory functions within a single device constitutes a critical approach to developing hardware systems characterized by homogeneity, high integration, superior compatibility and biocompatibility. Yet, an intrinsic trade-off exists between high-performance sensing capability (volatility) and compute-in-memory performance (non-volatility)—sensing functionality requires low ion transport energy barriers to facilitate rapid ion migration, whereas compute-in-memory functionality demands high ion transport energy barriers to enable long-term ion storage.
To overcome this trade-off and realize homogeneous multimodal sensory-memory-computing integrated hardware, a research team led by Professor Ma Wei from the State Key Laboratory for Mechanical Behavior of Materials at Xi’an Jiaotong University collaborated with Professor Wang Zhongrui’s group from the Department of Electrical and Electronic Engineering at the University of Hong Kong, alongside other partner institutions. They designed an organic electrochemical transistor (OECT) featuring a highly crystalline channel and a fully vertically interconnected structure. Based on ion-electron mixed conductor materials, this device is equipped with a channel that exhibits high crystallinity and a large aspect ratio. By modulating the magnitude of the gate voltage, selective control over amorphous/crystalline phase regions with distinct ion injection energy barriers is achieved. Concurrently, regulation of the gate interface processes enables the device to switch between volatile and non-volatile operational modes, demonstrating exceptional comprehensive performance (Figure 1).
As an artificial receptor, it boasts a high on/off ratio, excellent gain characteristics and a low subthreshold swing of 65 mV/dec, enabling real-time perception of diverse high-frequency weak electrophysiological signals. Furthermore, the device supports multimodal sensing of electrical, optical, thermal and chemical signals, substantially expanding the information dimension. As a non-volatile artificial synapse, it exhibits a number of conductance states (＞10⁵) far exceeding that of existing memristor technologies, coupled with ultra-low programming variability (△GDS²/σ²≈179). Across a 32-fold dynamic range, it maintains stable conductance states for over 10,000 seconds with minimal conductance drift. Finally, the researchers developed a homogeneous spiking neural network hardware and a reservoir computing system based on the aforementioned devices, demonstrating practical applications including image classification, conditioned reflex simulation and real-time cardiac disease diagnosis. This work holds profound significance for the advancement of miniaturized edge-computing AI systems.
Figure 1 (a) OECT serving dual roles as an artificial receptor and synapse for real-time information processing; (b) Sensing characteristics; (c) Compute-in-memory characteristics; (d) Homogeneous OECT array for constructing homogeneous neural networks; (e) Sensory-memory-computing integrated neural network applied in cardiac disease diagnosis.
The relevant research findings were published online in Nature Electronics under the title An organic electrochemical transistor for multi-modal sensing, memory and processing. Dr. Wang Shijie, a doctoral candidate from the School of Materials Science and Engineering at Xi’an Jiaotong University, and Dr. Chen Xi, a doctoral candidate from the Department of Electrical and Electronic Engineering at the University of Hong Kong, are listed as co-first authors. Professor Ma Wei and Professor Wang Zhongrui from the University of Hong Kong serve as co-corresponding authors. The State Key Laboratory for Mechanical Behavior of Materials at Xi’an Jiaotong University is designated as the first corresponding institution. Additional collaborators include Professor Ma En, Professor Zhang Wei and Professor Li Tao from Xi’an Jiaotong University, as well as Professor Li Yuxiang from Xi’an University of Science and Technology.
Simultaneously, Nature Electronics published a featured commentary titled An organic device with volatility on demand authored by Professor A. Alec Talin from Sandia National Laboratories, USA. The commentary notes that “the precisely tunable synergistic behavior of synaptic and sensing functions holds immense value for the development of multifunctional analog circuits that emulate the perception and learning capabilities of biological neural networks. Neural network arrays constructed using this device can minimize energy consumption and latency associated with analog-to-digital conversion and information transmission. Furthermore, the device’s intrinsic flexibility, environmental stability and cost-effectiveness underscore its promising application prospects in the field of brain-computer interfaces.”
Paper link:https://www.nature.com/articles/s41928-023-00950-y
Chinese
