Interface-Type (IT) memristors show great promise as next-genera-tion memory for neuromorphic computing. In this work, IT memristors based on \(\ce {Au}\)/\(\ce {Nb}\):\(\ce {STO}\) are fabricated, thoroughly tested, and the characterization is presented. However, IT memristors endure challenges such as low ON/OFF ratio and high nonlinearity, thereby reducing the accuracy and limiting their potential in neuromorphic systems. To address these issues, following optimization techniques are utilized: (i) Parallel connection of IT memristors to increase number of levels in composite device; and (ii) Implementing a nonlinear optimization to mitigate the inherent nonlinearity, which causes inaccurate weight updates. Simulation results indicate that the proposed approaches lead to improvements in the accuracy of the neuromorphic system: up to ∼ 18% in 2-segment models, and up to ∼ 59% in 3-segment models. Connecting memristors in parallel effectively achieves accuracy of ∼ 86% on MNIST dataset using a 2-layer fully connected neural network. Comprehensive analysis of results under different variations has been performed. The simulation results highlight the potential of IT memristors and optimization methods for advanced neuromorphic architectures.

Interface-type (IT) memristors show great promise as next-generation memory for neuromorphic computing. The IT memristors based on Au/Nb:STO are firstly fabricated and thoroughly tested, and the characterization is presented in this paper. Like other memristors, these IT memristors endure challenges such as low ON/OFF current ratio and high non-linearity, thereby reducing the inference accuracy and limiting their potential application in neuromorphic systems. To address these issues, the following optimization techniques are proposed: (i) Parallel connection of IT memristors to increase number of levels in composite device and suppress spatial variations; and (ii) Implementing a non-linear optimization to mitigate the inherent non-linearity, which causes inaccurate weight updates. Simulation results indicate that the proposed optimization approaches lead to a great improvement in the inference accuracy of the neuromorphic system: up to ~18% is observed in 2-segment models, and up to ~59% in 3-segment models. Connecting memristors in parallel effectively alleviated the impact of spatial variations, achieving accuracy of ~86% on MNIST dataset using a 2-layer fully connected neural network. Despite these advancements, the proposed optimization methods lead to increased energy consumption and consequently larger carbon footprint, thereby revealing an important trade-off between accuracy, overall performance, and carbon footprint. A linear increase in the carbon footprint has been observed with an increase in the number of parallel memristors, while 2-segment models have the lowest carbon footprint. Through a comprehensive analysis, the simulation results highlight the potential of IT memristors and optimization methods for advanced neuromorphic architectures.