Wafer-Scale Ferroelectric Transistors for Efficient Temporal Signal Processing by Delay-Coupled Physical Reservoir Computing

Recurrent neuron networks (RNNs) are well-suited for temporal signal processing owing to its recurrent connections. Amongst them, reservoir computing (RC) is of particular interest with fast and low-cost training. However, the long-term iteration of the reservoir costs considerable computation power, thus far hindering the application of RC. Realizing RC with physical nonlinear systems emerges as a promising solution to address this problem, where specifically designed physical nonlinear systems can be employed as the reservoir. Recent advances show that ferroelectric field effect transistors (FeFETs) with nonlinear switching and short-term memory are promising for the design and implementation of physical RC systems.<br/><br/>In this work, we present wafer-scale fabrication of MoS₂/BaTiO₃ FeFETs via a full solution process for the physical realization of RC. To develop the FeFETs, BaTiO₃ (BTO) film is first deposited on silicon wafer as the ferroelectric layer via a chemical solution deposition method. Specifically, a wafer-scale ~300 nm BTO film can be obtained after pyrolysis and crystallization. Large-area and continuous MoS₂ film is then deposited by spin-coating a MoS₂ dispersion on the ferroelectric BTO film. The dispersion contains few-nanometer thick MoS₂ nanosheets prepared by electrochemical exfoliation. Following this fabrication method, MoS₂/BaTiO₃ FeFETs with a short-term memory (memory fading time ~5 seconds) are developed. This short-term memory may originate from the retention degradation behavior of BTO as a result of the defects caused by the solution process. Besides, a memory window of ~10 V and an on/off ratio of ~10³ are demonstrated. With these suitable switching and memory characteristics for physical RC system realization, our FeFETs can be used to design the reservoir layer of physical RC systems. Particularly, a delay-coupled physical RC framework is designed in our work utilizing the memory effect of our MoS₂/BaTiO₃ FeFETs to simplify the reservoir layer. With this in mind, we design a hardware RC system for processing time series and temporal signals. Briefly, a raspberry pi is used for time-multiplexing, input, and output; ADC, DAC and TIA are integrated and used for communication between the raspberry pi and the analog reservoir layer implemented with our MoS₂/BaTiO₃ FeFETs. We prove that this hardware RC system after training can successfully forecast Mackey-Glass time series.<br/><br/>Given the scalability and low-cost of the MoS₂/BaTiO₃ FeFETs, we envisage our MoS₂/BaTiO₃ FeFETs based hardware RC system with further specific reservoir layer design and delay-coupled RC model training can achieve temporal signal processing for, for instance, pattern recognition from IoT sensor and healthcare data, and tracking dynamical motions in wearables and soft robotics.