Data-driven RRAM device models using Kriging interpolation.

Area of Science:

• Materials Science
• Electrical Engineering
• Data Science

Background:

• Resistive Random-Access Memory (RRAM) devices exhibit complex behavior and switching noise, making accurate modeling challenging.
• Traditional methods like binning may not fully capture the nuanced performance distributions of these devices.

Purpose of the Study:

• To propose and validate a novel two-tier Kriging interpolation approach for modeling RRAM jump tables.
• To accurately predict both the mean signal and the standard deviation of switching noise in electronic devices.

Main Methods:

• A two-tier Kriging interpolation model was developed, with the first tier predicting the mean and the second predicting the standard deviation.
• The approach was tested using 36 synthetic datasets with varying Gaussian distributions.
• Validation was performed on experimental data from TiOₓ devices, comparing predictions with actual distributions using statistical tests.

Main Results:

• The Kriging approach demonstrated superior accuracy in predicting both mean and standard deviation compared to conventional binning methods.
• The model successfully captured the complex behavior landscape and switching noise characteristic of RRAM devices.
• Kolmogorov-Smirnov and maximum mean discrepancy tests confirmed the model's predictive power on experimental data.

Conclusions:

• The proposed Kriging-based jump tables offer a more realistic modeling solution for RRAM and other non-volatile analog devices.
• This method can accurately represent device populations and analyze the impact of weight dispersion in neural network simulations.