12 January 2026 in Advanced Materials
9 May 2025 in Journal of Physics D: Applied Physics
Memristor-based analog in-memory learning (AIML) has emerged as a promising approach to improve energy efficiency in deep neural network training. However, non-idealities in memristive devices, such as nonlinearity, asymmetry, and cycle-to-cycle (C2C) and device-to-device (D2D) variations, pose significant challenges. These issues lead to increased energy consumption, reduced write precision, and compromised in-situ learning performance. To address these problems, we propose a mixed-precision training strategy that combines gradient accumulation with single pulse blind write method. We analyze the failure mechanisms of in-situ learning without these techniques and systematically investigate how various non-idealities affect AIML performance. We demonstrate that, by using our GA-Single Pulse strategy, high accuracy (95.36%) can be achieved even under significant non-idealities, including device conductance states being limited to 10 pulses for potentiation as well as 5 pulses for depression, the asymmetry of conductance state constrained to a factor of 2, the nonlinearity in long-term potentiation/ long-term depression curve reaching up to 5, C2C variation as high as 50%, and D2D variation extending up to 40% for learning handwritten digits in MNIST handwritten digit dataset, outperforms all previous reports. The results suggest that the idealities of memristive devices may not be as critical as previously assumed for AIML’s practical deployment.
8 July 2025 in Applied Physics Reviews
Memristor crossbar arrays, mimicking the human brain, hold immense potential for energy-efficient data-intensive computations in artificial intelligence applications such as image recognition and natural language processing. However, the stochastic nature of resistive switching (RS) in memristors often leads to poor device stability and uniformity, hindering the scalability required for real-world applications. Here, we present a novel phase segregation strategy to achieve uniformly distributed self-assembled Ni nanofilaments within a BaTiO3 matrix, enabling local deterministic redox reactions for RS, as confirmed by comprehensive structural, local, and macroscopic RS studies. This approach yields drastic enhancement in cyclic performance and device uniformity, with the average cyclic variances of Set voltage and low resistance state down to 1.4% and 9.6%, respectively. The devices also exhibit excellent endurance (109 cycles) and ultrafast programming speed (down to 100 ns) and achieve over 5-bit level long-term memory states. The enhanced cyclic stability and device uniformity translate to high training and learning accuracies (95%) in a three-level deep neural network, with 1-bit inputs. Our phase segregation strategy provides a generic pathway to overcome the long-standing challenge of device variability in neuromorphic computing.