Benchmarking Stochasticity behind Reproducibility: Denoising Strategies in Ta₂O₅ Memristors

Area of Science:

• Materials Science
• Neuroscience
• Electrical Engineering

Background:

• Memristive artificial synapses are crucial for neural network applications, requiring high reproducibility, endurance, and fine resolution of conductance weights.
• Inherent fluctuations in the active volume of memristors limit the resolution of synaptic weights.
• Understanding and mitigating noise is essential for advancing memristor technology.

Purpose of the Study:

• To introduce a comprehensive noise benchmarking and reduction protocol for memristive devices.
• To investigate the voltage-dependent noise characteristics and tunability in Ta2O5 memristors.
• To identify strategies for improving the signal-to-noise ratio in artificial synapses.

Main Methods:

• Developed a noise benchmarking protocol that tracks voltage-dependent noise along I(V) curves.
• Investigated noise level tunability using dedicated voltage cycling schemes in filamentary Ta2O5 memristors.
• Employed nonlinear noise spectroscopy to identify noise characteristics in different voltage regions.

Main Results:

• Identified significant order-of-magnitude variability in noise levels, even in seemingly reproducible switching cycles.
• Discovered a subthreshold voltage region exhibiting voltage-boosted fluctuations.
• Demonstrated that a few subthreshold cycles can reconfigure fluctuations, leading to a highly denoised state without resistive switching.

Conclusions:

• The developed protocol provides deep insights into memristor noise mechanisms.
• A specific subthreshold voltage manipulation strategy can effectively reduce noise in Ta2O5 memristors.
• This noise reduction technique holds promise for enhancing the performance of memristive artificial synapses in neural networks.