Applicability of the High Field Model: A Preliminary Numerical Study

In a companion presentation, we have discussed the theory of a mesoscopic/macroscopic model, which can be viewed as an augmented drift-diffusion model. Here, we describe how that model is used. The device we consider for this presentation is the one dimensional GaAs n+-n-n+ structure of length 0.8μm. First, a full hydrodynamic (HD) model, proven reliable when compared with Monte Carlo simulations, is used to simulate the device via the ENO finite difference method. As applied to the full device, the new model is not necessarily superior to traditional drift-diffusion (DD). Indeed, when we plot the quantity η = μ0E/ √ kT0 m , where μ0 is the mobility constant and E = −φ′ is the electric field, we verify that the high field assumption η > 1, required for the high field model, is satisfied only in an interval given approximately by [0.2, 0.5]. When we run both the DD model and the new high field model in this restricted interval, with boundary conditions of concentration n and potential φ provided by the HD results, we demonstrate that the new model outperforms the DD model. This indicates that the high field and DD models should be used only in parts of the device, connected by a transition kinetic regime. This will be a domain decomposition issue involving interface conditions and adequate numerical methods.


Introduction
In previous work, we have demonstrated the robustness of an algorithm (ENO: Essentially Non-Oscillatory) designed for the simulation of the hydrodynamic model for semiconductors over a wide range of parameters. In [5] and [6], n + -n-n + diodes in one dimension and MESFETS in two dimensions were simulated. The present paper deals with the high field model introduced in [3], and appearing elsewhere in these proceedings [2].

The Models
The new model is an extension of the drift-diffusion (DD) model. Hence we first describe the DD model in a context for comparison.

Drift-Diffusion Model in One Dimension with µ Depending on E
The DD model is well documented (see, for example, [4]). It is given by: where the representation in terms of hyperbolic and viscous components is given by J = J hyp + J vis , and J hyp = −µnE, Here, n denotes carrier concentration and J denotes current defined per unit charge modulus. We have separated current components in anticipation of the high field model to follow. Also, the electric field is denoted by E, so that in terms of the electrostatic potential φ, where we use the customary expressions, for doping n d , dielectric , and charge modulus e. Set τ denotes the relaxation time, m denotes effective mass and ambient temperature is denoted by T 0 , with θ given in energy units. Of course, µ is mobility, and we take µ to be dependent upon the electric field E, using the formula (2.42) in [4]: Here, v d has the interpretation of the saturation velocity and µ 0 is the low field mobility. The velocity for the DD model is a derived quantity computed by v = J n . The device we consider for this presentation is the one dimensional GaAs n + -n-n + structure of length 0.8µm. The device used is as follows: In the definition of µ, we take µ 0 = 4.0 µm 2 /V /ps. We consider T 0 = 300 K for which v d = 0.6 µm/ps, which is taken to be the maximum of the velocity in the HD run with vbias = 1.0.
The results are shown in Figure 1. They do not match as well with those of the HD model (with a doping n d dependent µ), as in the silicon case which is not presented here. The oscillations in the velocity near the left junction are due to the numerical differentiation of rapidly changing quantities. ********** place to insert Figure 1***********

The High Field Model
The model can be written as follows: where the representation in terms of hyperbolic and viscous components is now given by Also, the electrostatic equation is satisfied as in the DD model, and τ and θ are defined similarly.
Furthermore, ω is taken to be a constant: and the velocity is again a derived quantity.
We again consider the same GaAs device as before. The results are shown in Figure 2. Not much improvement, if any, is observed over the DD results. The spikes in the velocity near the junctions are due to the derived nature of that quantity; particularly the numerical differentiation of rapidly changing quantities.

Restriction to Channel
We plot the quantity η = µ 0 E/ kT 0 m , where µ 0 is the mobility constant and E = −φ is the electric field (obtained from the HD simulation), in

Comparison with Kinetic and Augmented Drift-Diffusion Models
The high field model may be thought of as a form of an augmented DD model. In this section, we briefly compare the results of our simulations with those of [1] and [7]. The results of [7] deal primarily with tracking the carrier drift velocity in Silicon. The model is very similar to that originally employed by Thornber [9]. We note that the perturbation terms of the high field model introduced here include more than simply differentiated electric field terms. Comparisons are made in [7] among various models, including DD and augmented DD, as well as the hydrodynamic model and Monte-Carlo simulation. In these studies, the authors of [7] tended to find that standard DD understated drift velocity, even in GaAs. In our own comparisons, when using derived velocity, and a field dependent mobility in DD, we did not experience such a pronounced understatement. We have presented the results of the simulations in Figure 1 for comparison with standard DD, and in Figure 2 for comparison with the high field model. And, remarkably, we find that in the channel, as a problem posed only there, DD overstates the derived velocity! Moreover, the high field model predicts a velocity intermediate between the DD model and the hydrodynamic model.
Paper [1] is difficult to correlate directly because emphasis is placed upon frequency distributions.
Nonetheless the average velocity and electric field calculations made there compare favorably with our own. Issues of heating and cooling discussed there are not raised in the present paper, but we intend to consider the effect of energetics in future work. The model has already been developed by the second author.

Algorithm
We shall only briefly describe the algorithm used in this paper, namely the ENO scheme developed in [8]. The ENO scheme is designed for a system of hyperbolic conservation laws of the form, where u = (u 1 , · · · , u m ) T , and the hyperbolicity condition, ∂f ∂u is diagonalizable, with real eigenvalues, holds. An initial condition is adjoined to (5.1).
For systems of conservation laws, local field by field decomposition is used, to resolve waves in different characteristic directions. For this purpose, analytical expressions are needed for the eigenvalues and eigenvectors of the Jacobian matrix. This reduces the determination of the scheme to the case of a single conservation law. Thus, to describe the schemes, consider the scalar one dimensional problem, and a conservative approximation of the spatial operator given by Here, the numerical fluxf is assumed consistent: The conservative scheme (5.2), which characterizes thef divided difference as an approximation to f (u) x , suggests thatf can be identified with an appropriate function h satisfying