Descriptive Predictive Modeling of Resistive Memory Retention
Ari Shtein, Noah Geisler, Dongjae Shin, Yiyang Li
Materials Science and Engineering, University of Michigan, Ann Arbor, MI
[1] L. Chua, “Memristor—The missing circuit element,” IEEE Transactions on Circuit Theory, vol. 18, no. 5, pp. 507–519, 1971, doi: https://doi.org/10.1109/tct.1971.1083337.
[2] Niu, Gang & Kim, Hee-Dong & Roelofs, Robin & Pérez, Eduardo & Schubert, Markus & Zaumseil, P. & Costina, I. & Wenger, Ch. (2016). Material insights of HfO2-based integrated 1-transistor-1-resistor resistive random access memory devices processed by batch atomic layer deposition. Scientific Reports. 6. 28155. 10.1038/srep28155.
[3] D. Shin et al., “Oxygen tracer diffusion in amorphous hafnia films for resistive memory,” Materials Horizons, vol. 11, no. 10, pp. 2372–2381, May 2024, doi: https://doi.org/10.1039/D3MH02113K.
This research was supported by the National Science Foundation through the Materials Research Science and Engineering Center at the University of Michigan.
Li+
Research Group
Li+
Research Group
Resistive Memory
y
HfO2 ReRAM
L. Chua introduced a fourth fundamental circuit element in 1971 which would relate charge q and flux-linkage Ѱ, called the memristor. He showed that it could ‘remember’ its resistance.[1]
Chua’s predictions for the basic behavior of his memristor
The 1T1R CMOS-integrated ReRAM architecture[2]
Li Group works to bring ReRAM—the computing memory implementation of the memristor—into industrial readiness.
A typical IV switching curve, annotated with device behavior
HfO2-based ReRAM is a focus of the Li Group.
Shin et al. recently explored physical explanations for why these devices retain memory so unexpectedly well.[3]
Setting the device creates a conductive ionic channel—pushing it into its low-resistance state. Then resetting dissolves that channel and puts it back in its high-resistance state
HfO2 interlayer ReRAM device schematic
Predictive Modeling
Artificial Reduced Layer
Future Work
Ta
More datasets, endurance, and learning to manufacture better devices.
Individual regressions with some features show very strong, statistically significant correlations
Instead, the retention of each device is predicted using a model trained only on the other devices in the dataset.
Predictions with one feature
(standard deviation set voltage)
Predictions with 11 features
(involving set & reset slopes, after-reset variation, and forming features)
With a threshold set around 8.5–9.5 hours of retention at 250° C, we can determine the model’s accuracy in predicting whether a device is truly non-volatile.
Reduced Layer
Reduced Layer
YSZ
Extracting Switching Features
Extracting Device Retentions
Each device is first formed and then switched 20 times.
The I-V data that results can be plotted, and ~30 features are extracted from the plot for each device.
Retention & Switching Features
For ReRAM to be an effective, industry-acceptable memory solution, it must be non-volatile. It needs to retain the resistive state it’s been set to for upwards of 10 years (and at temperatures up to 80-100° C).
After devices are set into their low-resistance state, they’re annealed at 250° C, and their resistances are read at specific intervals
The resistance readings are plotted against time, interpolated polynomially, and then my code looks for when its resistance becomes 5 times as high as the initial LRS.
All devices plotted
Example retention extraction