Multimodal ion-gated transistor based on 2D superionic conductor for in-memory computing in deep learning.

Abstract

For neuromorphic computing, integrating multiply-accumulate operations and nonlinear activation within a single device can reduce latency and power consumption while improving computational efficiency. However, these operations impose conflicting requirements: multiply-accumulate requires highly linear and non-volatile resistance states, while activation requires diverse nonlinear synaptic behaviors. These contrasting demands make integrating both functions in a single device challenging. Here, we construct a multimodal ion-gate transistor using 2D CdPS₃-Li as dielectric layer and MoS₂ as channel material. The layered structure of CdPS₃-Li facilitates anisotropic ion transport for Li⁺ storage and produces strong ion-electron coupling, resulting in high-linearity and non-volatile resistance states under electrical pulses. Moreover, Cd vacancies in CdPS₃-Li attract and trap photo-generated holes from MoS₂, leading to rich nonlinear behavior under light pulses. Therefore, the CdPS₃-Li transistor can simultaneously perform both operations. The CdPS₃-Li transistor arrays achieved high accuracy in handwritten digit classification, offering a promising hardware solution for neuromorphic computing.