Abstract
A common approach to model memristive systems is to include empirical window functions to describe edge effects and non-linearities in the change of the memristance. We demonstrate that under quite general conditions, each window function can be associated with a sigmoidal curve relating the normalised time-dependent memristance to the time integral of the input. Conversely, this explicit relation allows us to derive window functions suitable for the mesoscopic modelling of memristive systems from a variety of well-known sigmoidals. Such sigmoidal curves are defined in terms of measured variables and can thus be extracted from input and output signals of a device and then transformed to its corresponding window. We also introduce a new generalised window function that allows the flexible modelling of asymmetric edge effects in a simple manner.
Introduction
Memristors—resistors with "memory" whose resistance depends on past current—are critical components for next-generation computing. However, modeling their behavior remains challenging due to physical limits. As a device approaches its minimum ($R_{ON}$) or maximum ($R_{OFF}$) resistance, the rate of change in resistance slows down. Researchers typically use window functions, denoted as $f(z)$, to simulate these "edge effects" and ensure the internal state variable $z$ remains within the physical bounds $[0, 1]$.
Historically, these window functions have been chosen empirically or semi-empirically. This paper addresses the gap between these internal mathematical constructs and the observable electrical signals (voltage $v$ and current $i$) measured in experimental settings.
Bridging Window Functions and Sigmoidal Behavior
The core contribution of this work is the establishment of a formal mathematical bridge between internal window functions and the external sigmoidal response of a memristive system. The authors prove that for a standard class of models, the internal state $z$ follows a sigmoidal curve $S(q)$ relative to the injected charge $q(t) = \int_0^t i(\tau) d\tau$.
Specifically, the window function $f(z)$ is shown to be the derivative of the sigmoidal response with respect to charge. This bidirectional transformation allows researchers to derive accurate device models directly from measured data: if the saturation behavior of a device can be measured, its corresponding window function can be mathematically extracted using the relationship:
$$\frac{1}{f(z)} = \alpha \frac{d}{dz} [S^{-1}(z)]$$
The Generalized Window Function
To improve modeling flexibility, the authors propose a new generalized window function: $$f(z; r, p) = [\psi(r) z (1 - z^{r-1})]^p$$
This function introduces two key parameters that allow for more realistic device simulation:
- $r$ (Asymmetry): Controls whether edge effects are left-skewed or right-skewed. This is vital for modeling asymmetric ionic drift, which is common in real-world physical devices but absent in symmetric models like the original HP memristor.
- $p$ (Flatness): Controls the depth of the edge effects. Increasing $p$ creates a higher "charge threshold," meaning more integrated input is required before significant resistance changes occur.
Conclusion and Future Directions
By linking abstract window functions to observable sigmoidal curves, this framework provides a robust method for characterizing memristive devices. While the current model assumes the window function is independent of input polarity, it sets the stage for "piecewise ideal" modeling of non-ideal devices. Future work will involve applying these transformations to experimental data from RRAM and Phase Change Memory (PCM) to identify the sigmoidal "families" (such as Gompertz or Weibull) that best describe emerging hardware.
Summary updated