Tunneling‐Controlled Fusion of Short‐ and Long‐Term Memory in SiO2/HfO2‐Based Neuromorphic Device for Time‐Series Prediction

Based on trap‐assisted tunneling, the devices can fuse STM/LTM, where the low switching energy of 1 pJ and stable low‐power retention (0.2 % loss ratio and 3.05 × 10−11 W) is achieved. Training in a long short‐term memory network it allows to analysis time‐series data and then makes precise long‐term predictions with an error ratio of 4.465 %. Abstract Prediction is a needful action before decision‐making, typically related to the interaction of historical information and current states. It requires computing with information on multiple timescales, meaning intrinsic dynamics in short‐long‐term plasticity. However, current implementations of short‐long‐term memory are still limited by unstable retention in LTM mode and high energy consumption. Here a trap‐assisted tunneling strategy is used in Schottky‐based optical synapse, achieving the tunneling‐switched short‐long‐term memory transformation. Based on the devices, the spike‐timing‐dependent plasticity and spike‐number‐dependent plasticity have been programmed with a low switching energy of 1 pJ that is close to the human brain. By adjustment of light‐pulse inputs, the adaptive short‐to‐long‐term transition can be achieved with a clear mode jump. Furthermore, the formed LTM both in stability (only 0.2 % loss ratio over a span of 50 s) and energy requirement (as low as 3.05 × 10−11 W) shows desired improvement compared to other devices implementing short‐long‐term memory. By further mapping device parameters to a long short‐term memory network, it allows for the analysis time‐series data and then to make precise long‐term predictions with an error ratio of as low as 4.465 %. This research can offer a hopeful solution to address the precise trend prediction, holding substantial promise for intelligent applications. Based on trap-assisted tunneling, the devices can fuse STM/LTM, where the low switching energy of 1 pJ and stable low-power retention (0.2 % loss ratio and 3.05 × 10 −11 W) is achieved. Training in a long short-term memory network it allows to analysis time-series data and then makes precise long-term predictions with an error ratio of 4.465 %. Abstract Prediction is a needful action before decision-making, typically related to the interaction of historical information and current states. It requires computing with information on multiple timescales, meaning intrinsic dynamics in short-long-term plasticity. However, current implementations of short-long-term memory are still limited by unstable retention in LTM mode and high energy consumption. Here a trap-assisted tunneling strategy is used in Schottky-based optical synapse, achieving the tunneling-switched short-long-term memory transformation. Based on the devices, the spike-timing-dependent plasticity and spike-number-dependent plasticity have been programmed with a low switching energy of 1 pJ that is close to the human brain. By adjustment of light-pulse inputs, the adaptive short-to-long-term transition can be achieved with a clear mode jump. Furthermore, the formed LTM both in stability (only 0.2 % loss ratio over a span of 50 s) and energy requirement (as low as 3.05 × 10 −11 W) shows desired improvement compared to other devices implementing short-long-term memory. By further mapping device parameters to a long short-term memory network, it allows for the analysis time-series data and then to make precise long-term predictions with an error ratio of as low as 4.465 %. This research can offer a hopeful solution to address the precise trend prediction, holding substantial promise for intelligent applications. Advanced Science, EarlyView.