WorksPhysics and Astronomy
Modeling Defect-Level Switching for Highly-Nonlinear and Hysteretic Electronic Devices
Abstract
Previously, we demonstrated hysteretic and persistent changes of resistivity in two-terminal electronic devices based on charge trapping and detrapping at immobile metastable defects [H. Yin, A. Kumar, J.M. LeBeau, and R. Jaramillo, Phys. Rev. Applied 15, 014014 (2021)]; we termed these devices as defect-level switching (DLS) devices. DLS devices feature all-electronic resistive switching and thus are volatile because of the voltage-time dilemma. However, the dynamics of volatile resistive switches may be valuable for emerging applications such as selectors in crosspoint memory, and neuromorphic computing concepts. To design memory and computing circuits using these volatile resistive switches, accurate modeling is essential. In this work we develop an accurate and analytical model to describe the switching physics in DLS devices, based on the established theories of point defect metastability in Cu(In,Ga)Se2 (CIGS) and II-VI semiconductors. The analytical nature of our model allows for time-efficient simulations of dynamic behavior of DLS devices. We model the time durations of SET and RESET programming pulses, which can be exponentially shortened with respect to the pulse amplitude. We also demonstrate the concept of inverse design: given desired resistance states, the width and amplitude of the programming signal can be chosen accordingly.
Introduction
The trapping and detrapping of electronic charge at immobile point defects can change the conductivity of metal/insulator/metal (MIM) stacks. This was analyzed by Simmons and Verderber in 1967,[1] and their interpretation has been applied in many MIM material systems.[2, 3, 4, 5] This class of devices are known as all-electronic resistive switches. We recently proposed and demonstrated a mechanism for all-electronic resistive switching based on capture and release of electronic charge at metastable point defects at heterojunctions.[6] These metastable defects can exist in two distinct lattice configurations: a shallow- and a deep-double-donor state, and they will switch between these two configurations upon stimulation with light or voltage bias; we termed this effect as defect level switching (DLS). In the case of heterojunctions, the DLS-active defects are the sulfur vacancies in CdS.[7] Similar DLS phenomenology has been observed in other compound semiconductors, including DX centers in III-Vs,[8, 9, 10] and metastable defect complexes in the Cu-In-Ga-S-Se system (CIGS).[11, 12, 13, 14] In DLS devices, the state retention time is determined by point defect metastability; compared this to MIM stacks, in which the designed potential wells determine the retention time.
Because of the “voltage-time” dilemma analyzed by Schroeder et al., all-electronic resistive switching is not suitable for nonvolatile memory: the high current densities required for short READ/WRITE pulse (ns) will result in the retention time many orders of magnitude less than the benchmark value (~years).[15] However, for emerging applications such as selectors in a crosspoint array,[16] and neuromorphic computing concepts that operate in the temporal domain, such as reservoir computing,[17] the dynamical behavior of volatile resistive random access memory (RRAM) is valuable. For instance, Wang et al. utilized the dynamics of silver nanoparticles to functionally resemble the synaptic Ca behavior.[18] Using the volatile switching effect of silver filaments, Zhang et al. achieved the emulation of multiple essential features of artificial neuron in a unified manner without auxiliary circuits.[19] Quantitative modeling of the dynamics of conductive filament-based RRAM devices has been extensively studied, far more so than that for all-electronic RRAM.[20, 21, 22, 23, 24, 25] All-electronic volatile RRAM may prove particularly useful if the switching dynamics can be predictively modelled based on established semiconductor physics concepts, and if devices not relying on mass transport prove more reliable in operation than devices relying on ion motion, nanoscale redox chemistry, and other complex coupled electronic-ionic phenomena. In this context, DLS devices occupy a unique niche: they feature dynamics that can be quantitatively modeled and tuned through materials selection and device design, and they operate on the basis of charge state transitions at immobile defects.
Here we develop an analytical and highly efficient model to describe the dynamics of DLS devices, starting from the statistics of point defect charge state transition. In § 2, we describe the statistics of DLS-active point defects, following established theories of metastable point defects in CIGS solar cells,[11, 12, 13, 14] and II-VI semiconductors.[7, 26, 27] In § 3, we model a metal-semiconductor heterojunction with DLS-active defects, as we have demonstrated experimentally.[6] The resistive switching is caused by DLS defects that remain in metastable charge states after voltage stimulus, resulting in an out-of-equilibrium band diagram. Using Kimerling’s model of deep traps, we derive an analytical expression of the interface band diagram by approximating the space charge profile using a step function.[28] The effects of bias treatment on the junction conductivity are consistent with our experimental results, and the heterojunction band diagram and metastable defect distributions computed from our model are in great agreement with those computed from numerical iteration (as proposed by Decock et al.).[14] Also, we estimate that our model is at least one million times more efficient compared to the numerical approach.
The analytical nature of our model allows for simulations of the device dynamics in a time-efficient manner. In § 4, we illustrate the nonlinear and hysteretic resistive switching of a DLS heterojunction device upon a sinusoidal driving voltage. We model and discuss the hysteretic behavior at different frequencies of the driving voltage. We model the time durations of SET and RESET programming pulses, which can be exponentially shortened with respect to the pulse amplitude. These simulations demonstrate the concept of inverse design: given desired resistance states, the time width and voltage amplitude of the programming signal can be chosen accordingly. The concept of inverse design may extend to materials selection led by circuit design and system optimization, provided that a larger and accurate database of DLS-active defects becomes available.
Configuration transition kinetics of DLS-active defects
DLS-active defects are found in a number of compound semiconductors, including DX centers in -AlGaAs and III-nitrides,[8, 9, 10] the oxygen vacancy in -ZnO,[26, 27] the sulfur vacancy in -CdS.[7] Though in this section we use specific features of anion vacancies in wide-bandgap II-VI semiconductors to discuss configuration transition kinetics, the model can be generalized to any other semiconductor system containing DLS-active defects.
In Figure 1 we present the configuration coordinate diagram for DLS-active defects, using for illustrative example the case of anion vacancies that are double donors. We represent the deep, neutral state of the anion vacancy as , which is occupied by two electrons, corresponding to in ZnO and in CdS; upon ionization, can transform into the doubly-ionized, shallow donor state ; the asterisk indicates a metastable lattice configuration. In ZnO, due to the negative- effects the singly-ionized state is never thermodynamically stable.[26, 29] In §A1 we plot the equilibrium distributions of different charge states for the oxygen vacancy, and the concentration of singly-ionized state is at least ten orders of magnitude smaller than or . Therefore, the transitions and are experimentally inaccessible, and we only address the configuration conversion between the neutral and doubly-ionized metastable states and .

Figure 1. Configuration coordinate diagram for DLS-active point defects that exhibit metastable behavior – here, anion vacancies that are double donors. and represent stable configuration coordinates for the neutral deep donor state and doubly-ionized shallow donor state , respectively; the asterisk indicates a metastable lattice configuration. System enthalpy vs. configuration coordinate and are represented by orange and green solid lines, respectively. Thermal activation energies and necessary for transitions in Eq. (1) and Eq. (4) are indicated by orange and green arrows, respectively.
In -type semiconductors, the transition from deep to shallow donor configuration can happen via electron emission process, and the reaction path is given by:
The transition in Eq. (1) requires simultaneous emission of two electrons as well as thermal activation over an energy barrier , as indicated in the configuration coordinate diagram in
Fig. 1. Accompanying the deep-to-shallow configuration transition, the cation-cation distances around the anion vacancy may increase by upwards of 30%.[30] An energy of is required to distort the lattice configuration, and shift the defect level to be resonant with the conduction band. The optical transition energy is higher than the thermal energy barrier, because crystal-momentum-conserving optical transitions can occur only at fixed configuration coordinate.[8] In addition to the reaction path described in Eq. (1), when valence band holes are injected through bias treatment in a junction, the deep-to-shallow transition is possible via simultaneous capture of one hole and emission of one electron, or simultaneous capture of two holes, with thermal energy barriers and , respectively, and the corresponding reaction pathways are given by:
As shown in Figure 1, the reverse transition from the metastable state to the neutral state happens through simultaneous capture of two electrons from the conduction band as well as thermal activation over an energy barrier , and the reaction path is given by:
In principle, the transition from the doubly-ionized state to the neutral state can happen through hole emission processes, i.e., simultaneous capture of one electron and emission of one hole, or simultaneous emission of two holes. To emit holes to the valence band, the defect needs to be in acceptor states. Because the defects discussed in this work (anion vacancies in II-VI semiconductors) are either in shallow donor states or deep donor states, hole emission processes are unlikely to happen. However, in some other semiconductor systems, the contribution from hole emission needs to accounted for. For instance, the defect complexes in CIGS are amphoteric and can transform between donor and acceptor states.[13] In this case, the donor-to-acceptor configuration transition pathways should include those through hole emission processes.
The energy barriers related to these configuration transitions can be calculated by ab initio methods or inferred from experimental data. Using the local density approximation (LDA) of density function theory (DFT), Lany and Zunger calculated the configuration coordinate diagram for the oxygen vacancy in ZnO, which gives , and .[27] For in CdS, using the experimental data on photoconductivity, we infer that the thermal activation energy for electron recombination is ; using the experimental data on resistivity of MoO/CdS heterojunction devices, we infer .[6, 7] In the presence of free holes, the deep donor DX state is immediately ionized without thermal activation, i.e., for both defects.[31] An important future work for us is to use DFT calculations to accurately compute the configuration coordinate diagram for as well as other DLS-active defects, and compare the calculations results to the experimentally determined activation energy barriers.
Now we discuss the kinetics of transitions described in Eqs. (1)–(4). According to the Shockley-Read-Hall (SRH) mechanism, we denote one-electron emission/capture rates as , which are given by and , respectively, where is the carrier thermal velocity, the effective density of states of the conduction band, the electron density in the conduction band and the carrier capture cross section. Since is given by , where is the energy level of the conduction band and the Fermi level, we can rewrite the electron capture rate as . We denote the transition rates of Eqs. (1)–(4) as , , and , respectively. For the transition in Eq. (1), the attempt frequency for activation over the energy barrier is comparable to the phonon frequency , and the probability of one successful activation attempt is . Following Lany and Zunger’s analysis, after one successful activation attempt, the time window for electron emission is , and during this time window the probability of two-electron emission is . [13] Hence, the transition rate is given by the following expression:
Here, the assumption is that the one-electron capture cross sections are the same in the first and second ionization processes. We acknowledge that it’s a crude assumption, and it may lower the accuracy of simulation results. According to the analysis by Chicot et al., the cross section is larger in the second ionization than the first one, based on an argument of the Huang and Rhys parameter.[32] However, more rigorous analysis and calculations are required to accurately determine the capture cross sections in the two ionization processes. As we will show in § 4, Eq. (25) determines the device dynamics and the capture cross sections only enter the prefactor of Eq. (25), without affecting the hysteretic behavior and voltage dependence and of the device dynamics. Therefore, in this work we are not bothered unduly by this assumption.
Following our discussion on , , and are given by:
where is the effective density of states of the valence band.
We should note that in this work the charge transition rates are basically calculated by the SRH model, and we account for the lattice configuration changes by introducing Boltzmann factors. The SRH theory is a simplified model and it was shown to be invalid to describe the charge exchange rates between two materials.[33] In the latest works, advanced defect models such as the non-radiative multi-phonon (NMP) theory and the Marcus charge transfer theory combined with DFT calculations have been used to accurately calculate the charge transition rates.[33, 34, 35] [36] Integration of these advanced defect models into our model will effectively improve the accuracy of simulation results, that we leave in our future work.
We denote the shallow-to-deep and deep-to-shallow configuration transition rates as and , respectively. For the shallow-to-deep transition, Eq. (4) is the only pathway considered in this work. For the deep-to-shallow transition, the two-hole capture process in Eq. (3) is the dominant mechanism because it has the lowest thermal activation energy. The other processes have higher energy barriers and can gain importance at high temperatures. In §A1 we analyze , and as a function of Fermi energy and demonstrate that is the most important contribution upon junction reverse biasing and hole injection. Therefore, and , which also preserves the symmetry of rate equations:
We denote the Fermi level position where the deep donor state and the metastable shallow donor state are equally likely as the transition energy , i.e., when . Using Eq. (9) and Eq. (10), we have:
where is the semiconductor band gap. For , and the deep donor state is thermodynamically favored; for , and the metastable shallow donor state is thermodynamically favored.
If we denote the concentration of DLS-active defects as , and the concentration of deep and shallow donor states as and , respectively, the time variation of and are given by:
At steady states, the time variation of and are both equal to zero, yielding steady state concentrations and :
The equations in this section connect the dynamics of DLS devices to defect and material parameters, via fundamental semiconductor statistics. In Table 1, we list the properties of DLS-active defects in II-VI semiconductors, i.e., in ZnO and in CdS, and other material parameters, which we will use in simulations in § 3 and § 4. For future work, a computational database of DLS-active defects is required. The database should include defect parameters which are necessary to guide the design of DLS devices, such as thermal activation energies of configuration transitions. These properties could differ greatly from one defect to another, and should computed carefully and accurately by DFT calculations and calibrated by experiments. Such a computational database, combined with the model that we develop in this work, suggests promising opportunities for technology computer-aided design (TCAD) of resistive switching devices based on DLS phenomena.
Resistive switching physics in DLS-heterojunction devices
Configuration conversion between deep neutral and ionized states of DLS-active defects leads to persistent photoconductivity, and is useful in high-responsivity photoconductive sensors. However, it can also be achieved without illumination by applied bias in a metal-semiconductor Schottky junction, where the metal serves as a hole injection layer to induce the configuration transition defined in Eq. (3). Though operating at an interface, DLS devices are not based on the valence-change resistive switching mechanism in which crystal defects drift under an electric field, and there is no mass transport in DLS devices.37 In case of the MoO/CdS heterojunction that we have experimentally demonstrated, MoO has a higher work function than CdS and thus can inject holes into CdS, leading to a space charge layer (SCL) within CdS.[6] Within the SCL, most metastable defects are in the shallow donor state (). In Fig. 2a we present the energy levels of materials used to make a DLS-heterojunction device: an -type semiconductor with DLS-active defects, and a metal or highly-doped semiconductor. and refer to the metal work function and the semiconductor electron affinity, respectively. In the semiconductor, and refer to the energy levels of deep and shallow donor states, respectively. Because of the negative- effects, the first ionization energy (corresponding to the transition) is deeper than (corresponding to the transition).38 The Fermi level is above the transition energy . We denote by the energy difference between and the Fermi level . In Fig. 2b we present the schematic band diagram of an unbiased DLS-heterojunction. The formation of Schottky barrier is due to the metal work function being larger than the semiconductor electron affinity . In the semiconductor side, the conduction band edge energy and the metastable state transition energy vary with distance from the heterojunction. We denote by the location where intersects the Fermi energy:
In Fig. 2b the position of is indicated by a black star. According to Eqs. (14)–(16), is where both of the steady state concentrations and are equal to , i.e., and are equally likely at . Because the metastable state distribution affect the band bending and , and have to be calculated iteratively for self-consistency. The calculation procedure is as described by Decock et al. in the context of studying apparent doping profiles in CIGS solar cells that contain metastable – defect complexes.[14]
defines where and are equally likely at equilibrium. Under time-varying external stimulus, the distribution of metastable states may be driven out-of-equilibrium, and may no longer describe the position where and are equal to . We define as the instantaneous position where the deep and shallow donor state concentrations are the same:
At , ; while at , . At thermodynamic equilibrium, .
Under positive bias applied to the metal side of the heterojunction in Figure 2b, the concentration of electrons in the space charge region of the semiconductor will increase. Consequently, around where and coexist, electron recombination will convert to as described in Eq. (4). As a result, fewer defects are in the ionized state and more defects are in the deep neutral state , and will move closer to according to Eq. (17). This process requires thermal activation, and it’s thus slow to follow the applied bias, leading to hysteretic redistribution of and . Hence, , and are dynamical, history-dependent variables.

Figure 2. Figure 2: Schematic of a DLS-heterojunction. (a) Illustration of energy levels of the materials used: an -type semiconductor with DLS-active defects, and a metal or highly-doped semiconductor. Terms are described in the text. (b) Illustration of the energy band diagram after junction formation. The black arrow indicates the height of the Schottky barrier . The gray dash-dotted line indicates the location of the interface where . The black star represents , i.e., the intersection of the transition energy and Fermi level . For , and the shallow donor state is thermodynamically favored; while for , and the deep donor state is thermodynamically favored.
To track the evolution of and , the Poisson equation must be solved at each time to compute and , which are used to calculate the configuration transition rates defined in Eq. (12) and Eq. (13), and then update and at next time step. Solving the Poisson equation at each time is computationally demanding because it requires numerical iteration. For faster computational simulation of the device dynamics, we simplify the profile of space charge density to a step function, as was proposed by Kimerling for junctions that contain deep traps.[28] Kimerling assumed that deep traps contribute to the space charge only within a region of width from the interface, which bears the same meaning as the in Eq. (17). Hence, the space charge density is given by the following expression:
For the cases simulated here, DLS-active defects are chosen to be the dominant type of defects in the materials discussed, i.e., is much larger than the shallow (non-DLS) donor concentration (Table 1), and the contribution from regular shallow donors to the space charge profile is negligible. For cases in which both metastable and shallow defects make meaningful contributions to the space charge profile, as in CIGS thin films, Decock et al. proposed to use a staircase space charge density profile (this multi-step model can be seen as an extension of the one-step model).[14] In §A2 we explicitly compare the results of this staircase approximate model to direct numerical simulation and find good agreement. Therefore, the assumption of step-function changes in space charge density profile is generally applicable to junctions with deep and shallow levels.
| Material / Parameter | ||
| (eV) | 2.4 | 3.4 |
| (V) | 4.3 | 4.5 |
| 9 | 11 | |
| (cm) | ||
| (cm/s) | ||
| (cm/()) | 2 | 200 |
| DLS-active defects | Sulfur vacancy | Oxygen vacancy |
| (eV) | 1.1 | 1.7 |
| (eV) | 0.6 | 0.2 |
| (eV) | 0.0 | 0.0 |
| (cm) | ||
| (THz) | 1 | 1 |
| (cm) | ||
| (cm) | ||
| Thickness (um) | 0.06 |
Table 1. Properties of two wide-band gap II-VI semiconductors that host DLS-active defects: in ZnO and in CdS. is calculated according to Eq. (11). For the carrier capture cross section , Chicot et al. identified a deep donor of a large electron capture cross section of cm to be related to in ZnO.[32] The order of magnitude of cm is in agreement with a positively charged defect center and we assign cm for both in CdS and in ZnO. We take the phonon frequency THz for both defects. and are the concentrations of DLS-active defects and non-DLS-active shallow donors. We keep the same as the metastable defect concentration that we use in device simulations in our previous work.[6] The layer thickness is used in the simulations below.
We use Gauss’s law to calculate the electric potential profile variation through the junction. The boundary conditions of the conduction band energy (referenced to the Fermi energy) on the two sides of the semiconductor are:
and are the dielectric constant and width of the semiconductor, respectively. is the applied bias at the given time . is then given by:
where is given by:
Taking the parameters of the MoO/CdS heterojunction (that we list in Table 1, and for MoO we take its work function to be 6.7 eV), we illustrate the effect of bias on the interface band diagram and metastable defect states in Figure 3. As shown in Figure 3a and Figure 3b are the transition energy level and the concentration of shallow donor states in the CdS layer. is computed by combining Eq. (11) and Eq. (20) while is computed from Eq. (15). In Figure 3: the dashed blue line indicates or at zero bias, while the orange and green solid lines indicate those after forward bias treatment (+0.8 V) and reverse bias treatment (-0.8 V), respectively; in Figure 3b, the vertical dashed lines indicate positions of . After forward bias treatment, moves from 16.1 nm to 9.5 nm. According to Eq. (17), the decrease of means fewer defects are in the ionized shallow state, which reduces the effective donor doping concentration and the electric field at the junction; as a result, the interface barrier is thicker than before, switching the junction to high resistance state (HRS). After reverse bias treatment, moves from 16.1 nm to 20.7 nm, meaning that more defects are in the ionized state, increasing the effective donor doping concentration and the electric field at the junction; therefore, the interface barrier is thinner than before, switching the junction to low resistance state (LRS). These simulation results are consistent with the bipolar resistance switching that we demonstrated experimentally: DLS devices switch to HRS under forward bias, and switch to LRS under reverse bias.[6] We note that the changes of defect charge states and depletion region upon bias treatment shown in Figure 3 are in analogy to those in deep level transient spectroscopy (DLTS) measurements. In DLTS, a forward bias treatment can neutralize defects and widen the depletion region, causing a temporary decrease of the junction capacitance; in this work, the thickening of interface barrier induced by forward bias treatment causes a decrease of the junction conductivity.

Figure 3. Relationship between interface band diagram and metastable defect states. (a) Transition energy level that is computed by combining Eq. (11) and Eq. (20): without bias treatment (blue dashed line), after forward bias (orange solid line) and reverse bias treatments (green solid line). The square symbols represent the results that we compute using the numerical iteration method developed by Decock et al.[14] Energy is referenced to the Fermi level (grey dot-dash line). (b) Distribution of shallow donor states that is computed from Eq. (15): without bias treatment (blue dashed line), after forward bias (orange solid line) and reverse bias treatments (green solid line). The vertical dashed lines indicate positions of . The heterojunction is in HRS and LRS after forward and reverse bias treatments, respectively, and the corresponding positions of are labeled as and , respectively.
To examine the accuracy and efficiency of our model, we use the numerical iteration method developed by Decock et al. to compute and .[14] The results are shown as square symbols in Figure 3 and they are in good agreement with those from our model which are represented by dashed and solid lines. Also, we estimate that our analytical model is at least one million times more efficient than the numerical iteration approach: to compute one band diagram in Figure 3a, it takes less than 60 micro-seconds with Eq. (20), while it takes about 60 seconds to achieve the self-consistency with the numerical iteration method. All our simulations are executed with Python code and NumPy package.
Using the interface band diagram that we compute from Eq. (20), we estimate that after forward bias treatment the junction resistivity is 783 times that under zero bias (conductivity modeling described in §A3). At the doping concentration , as in Table 1, tunneling is the dominant transport mechanism, and it is exponentially dependent on both of the height and width of the interface barrier. In experiments, we found that the MoO/CdS heterojunction resistance could increase by up to 200 times after a forward bias treatment of +0.8 V at room temperature.[6] When the concentration of DLS-active defects is reduced to , we calculate the resistivity ratio of the forward-biased to the zero-biased junction to be 187, which is close to the experimental value.
As shown in Figure 3b, in the HRS more metastable defects are in the neutral deep state (more electrons are trapped at the interface) while in the LRS more are in the ionized shallow donor configuration (fewer electrons are trapped at the interface). The physical insight is the same as presented by Simmons and Verderber: trapped charge distorts the interface band structure, and as a consequence changes the electric field and tunneling currents passing through the interface.[1]
Eq. (20) and Figure 3 indicates that solely determines the interface band diagram as well as the resistivity. Therefore, is a state variable of the DLS device and the device dynamics are represented by the movement of . Its velocity, , can be determined by the space and time variation of at :
Combining Eqs. (9)-(12) and (17), and estimating , we have the numerator of Eq. (22) given by the following expression:
Given the boundary conditions of in Eq. (19), we have , which indicates that the numerator of Eq. (22) can be increased exponentially under applied bias.
We rewrite the denominator of Eq. (22) as:
represents the characteristic length of the truncation of the shallow donor distribution around . For the shown in Figure 3b, we calculate to be 3.1 nm for the junction under zero bias, and 3.6 nm and 2.8 nm for that after forward and reverse bias treatment, respectively. Since only changes slightly after bias treatment, we take it to be constant and thus Eq. (24) is constant. Combining Eqs. (22)–(24) yields:
where is a constant velocity, that is given by:
In deriving Eq. (25), we have simplified the complex process of defect ionization and electron recapture in DLS devices into the motion of , while preserving the essential switching physics. The expression of the velocity of has a similar form to that describing the growth velocity of conductive filaments in metal oxide resistive switching devices: both have a hyperbolic-sinusoidal dependence on the applied bias, which means that in both cases the resistive switching is bipolar and the switching speed can be increased exponentially with the drive voltage amplitude.39 However, the switching physics is different: in the case of DLS-heterojunction devices, the exponential dependence stems from electron or hole injection whose densities are exponentially dependent upon applied bias, while in the case of metal oxide resistive switching devices, the exponential dependence originates from electric field driven ion hopping. Also, in Eq. (25) all of the parameters are constructed from point defect physics, and there is no mass transport.
Dynamic behavior of DLS-heterojunction devices upon voltage stimulus
As an example, in Figure 4a, we show the response of (red solid line) upon a sinusoidal driving voltage (blue solid line) at a frequency of 1 MHz and an amplitude of 1.8 V. We also show the steady-state (i.e., equilibrium) position of upon the driving voltage with the gray dashed line, and we see that the device operates in out-of-equilibrium states of . As illustrated in Figure 4b, is hysteretic with respect to the driving voltage, and it changes in the interval of 13.8–17.4 nm. When the device is at HRS, nm, and when the device is at LRS, nm. Under this sinusoidal driving voltage, we estimate the resistivity ratio of HRS to LRS to be 50.

Figure 4. Resistive switching and hysteresis behavior of a DLS-heterojunction device under a sinusoidal driving voltage whose frequency is 1 MHz. In (a) we plot the voltage stimulus (blue solid line) and the corresponding change of (red solid line) versus time. We also plot the equilibrium position of (gray dashed line) versus time. The device operates in out-of-equilibrium states of . In (b) we plot with respect to the voltage stimulus. We label the processes of SET (from HRS to LRS) and RESET (from LRS to HRS) with black arrows.
Since the hysteretic resistive changes are based on non-equilibrium charge states of metastable point defects, DLS devices are predicted to be volatile and not suitable for nonvolatile memories.[15] At lower driving frequency, we expect reduced resistive switching because the metastable defects have more time to approach their thermodynamic equilibrium states. For sufficiently low frequency (smaller than the order of magnitude of the inverse of the device retention time), the hysteretic loop shown in Figure 4 will collapses into a line, i.e., the hysteresis vanishes and closely follows the applied bias. The room-temperature retention time of the HRS at MoO/CdS heterojunctions has been found to be about 90 s.[6] At the HRS, (as previously demonstrated in Figure 3b), and the DLS-active defects will gradually recover their thermodynamic equilibrium states, by emptying the occupied electrons through electron emission. Using Eq. (5) and the experimental fact that s, we estimate eV. impacts the device dynamics at low driving frequencies. Adding Eq. (5) into Eq. (10) and Eq. (12), we can study the hysteretic behavior of at different frequencies and compute the corresponding resistivity ratio of HRS to LRS. In Figure 5 we show the resistivity ratio upon sinusoidal driving voltage of different frequencies; insets show hysteretic loops of . At frequency Hz, which is much lower than the inverse of the retention time, i.e., s, the hysteretic loop of collapses into a line as expected, and . With increasing frequency, increases to a maximum and then decreases. The decrease is because, at very high frequency, defects have insufficient time to switch between configurations. As indicated by Eq. (25), the equation of motion of , the resistivity ratio is limited by the single electron/hole emission rates, the configuration transition barriers, the driving voltage amplitude, as well as other material properties. Using higher voltage amplitude will increase the injected carrier concentration and thus increase the resistivity ratio .

Figure 5. The resistivity ratio of HRS to LRS upon sinusoidal driving voltage of frequencies from 0.1 mHz to 1 MHz. In the three insets we plot the hysteretic loop of (i.e., the change of with respect to the voltage stimulus) at three different frequencies: 0.1 mHz, 10 mHz and 1 MHz.
Using Eq. (25), we can study the switching time between HRS and LRS of DLS-devices under voltage pulses, which are typically used to program resistive states in RRAM devices, and are relevant to neuromorphic computing strategies. To understand the switching process intuitively, we derive analytical expressions of the switching time. Eq. (25) is analytically solvable if we expand linearly with respect to , yielding:
where is the equilibrium position of at zero bias (as previously demonstrated in Figure 3b). Eq. (27) approximates fairly well in the range of and that we are interested in (§A3).
Under reverse voltage pulse , the device changes from HRS to LRS (the SET operation): moves from to . Combining Eq. (25) and Eq. (27), we solve for the switching time :
Under forward voltage pulse , the device changes from LRS to HRS (the RESET operation): moves from to . Combining Eq. (25) and Eq. (27), we solve for the switching time :
We see in Eq. (28) and Eq. (29) that the SET/RESET switching times can be exponentially reduced by increasing the pulse amplitude . So, using voltage pulses of larger amplitude is preferred to achieve faster switching. We note that is only dependent on and independent of ; vice versa, is solely dependent on , regardless of the position of .
Eq. (28) and Eq. (29) enable predictive device design and selection of operating parameters. Provided that we are required to design a device with a resistance ratio of 50. From Figure 4 we learn that using a combination of nm and nm can achieve a resistance ratio of 50. Then, to switch the between 13.8 nm and 17.4 nm, we need to compute the time widths ( and ) and amplitude () of programming pulses.
In Figure 6a we compute and plot and as a function of , using Eq. (28) and Eq. (29); the pulse time widths and amplitude should be chosen accordingly. For instance, we use a pulse amplitude of V; then the pulse time widths need to be ns and ns, respectively. In Figure 6b we demonstrate the dynamic evolution of under a train of such SET and RESET voltage pulses, and the starts from the equilibrium position , and switches between and as designed. Here, the evolution of is computed with Eq. (25).

Figure 6. SET/RESET of a DLS-heterojunction device with a resistance ratio of 50, triggered by voltage pulses, and of HRS and LRS are set to be 13.8 nm and 17.4 nm, respectively, i.e., nm and nm. (a) Time widths of voltage pulse required to trigger the SET (green solid line) and RESET (orange solid line) versus amplitudes of voltage pulse. (b) A train of SET and RESET voltage pulses (blue solid line) that switch the device between HRS and LRS. The voltage pulses have
an amplitude of and time widths of and , respectively. The grey dashed line indicates the position where the voltage is zero. We compute and plot the dynamic evolution of (red dashed line) using Eq. (25). starts from , which is the equilibrium position of at zero voltage, and it then switches between and as designed.
Discussion and conclusion
In this work, we use the statistics of metastable point defects and an approximation of step-changes in charge density to develop a physics-based model of resistive switching dynamics in DLS-based heterojunction devices, as we previously demonstrated experimentally.[6] The analytical nature of our model allows for highly efficient device simulations, compared to iterative methods, and we validate our model approximations by reference to the numerical iteration. We explore how devices respond to voltage pulses and variable-frequency drive.
DLS devices are based on charge trapping/detrapping at immobile point defects, and thus the relevant resistive switching mechanism is purely electronic and inherently volatile. Therefore, in DLS devices the programmed resistive state will relax into a thermodynamically stable state after removing the programming signal, but the relaxation may offer desirable dynamics for the emulation of biological neurons and synapses.[24] A notable feature of DLS devices is that they don’t suffer from stochastic ion migration events which occur in devices based on mass transport, such as conductive filament (CF) based RRAM.40 As a result, DLS devices are more amenable to physics-based design and are likely less affected by cycle-to-cycle variation and offer long-term stability.
Predictive modeling of these volatile resistive switching devices will help in designing computing circuits that leverage their dynamics. For future work, a computational database of DLS-active defects is required. The database should include defect parameters which are necessary to guide the design of DLS devices, and these parameters can be computed by DFT calculations and calibrated by experiments. Such a computational database of defects, combined with the physics-based model presented here, suggests a future in which circuits employing DLS devices could be optimized through an inverse design process: first by identifying the circuit dynamics that best suit the proposed algorithm, then resolving dynamics to the level of individual devices, and then achieving the desired device dynamics by materials selection. Volatile DLS devices with designer dynamics could be useful for circuits of coupled, hysteretic, and nonlinear oscillators, such as concepts in physical reservoir computing.41 Our model is also useful in other junction devices containing metastable defects, such as solar cells, in which metastable defects have direct impact on device performance.42
Acknowledgments
This work was supported in part by the Office of Naval Research MURI through Grant No. N00014-17-1-2661, and in part by the Advanced Concepts Committee of MIT Lincoln Laboratory.
Appendices
Equilibrium distributions of different charge states and configuration transition rates for the oxygen vacancy in ZnO
The singly-ionized state is thermodynamically unstable, i.e., the energy of the singly-ionized defect is always higher than the neutral state and the metastable state , as shown in Figure A1a[26, 27]. According to the formation energies, we compute the equilibrium concentrations of different charge states of the oxygen vacancy as a function of Fermi energy, as shown in Figure A1b. The dominant charge state is either or , and the concentration of the singly-ionized state is at least 10 orders of magnitude smaller than the dominant charge state. Therefore, the transitions and are experimentally inaccessible, and we only consider and model the transition rates between neutral state and doubly-ionized metastable state in this work.

Figure A1. (a) Formation energies and (b) equilibrium concentrations of charge states , and for the oxygen vacancy in ZnO as a function of Fermi energy at temperature T=300 K. is the valence band energy level. Formation energies were calculated by Lany and Zunger.[27] Here, the total concentration of oxygen vacancies is . We use black solid line, black dashed line and grey dash-dotted line to represent the formation energy or equilibrium concentration of charge state , and , respectively.
We plot the configuration transition rates (green solid line, defined in Eq. (8)), (grey solid line, defined in Eq. (7)), (blue solid line, defined in Eq. (5)) and (orange solid line, defined in Eq. (6)) as a function of Fermi energy at temperature T=300 K and T=1000 K in Figure A2a and Figure A2b, respectively. , and can contribute to the deep-to-shallow transition rate . As shown in Figure A2, at a high concentration of injected holes, is the dominant contribution. Also, the configuration transition rates need to be comparable to the frequency of driving voltage (i.e., , which is indicated by a black dashed line in Figure A2) to have an impact on the device dynamics. At T=300 K, is way lower than .

Figure A2. Configuration transition rates (green solid line), (grey solid line), (blue solid line) and (orange solid line) as a function of Fermi energy at temperature (a) T=300 K and (b) T=1000 K, respectively. is the valence band
energy level. We use a black dashed line to indicate the frequency of driving voltage that we use in device simulations, i.e., 1 MHz = .
Calculation of apparent defect density profiles using staircase space charge density profile
We reexamine the calculation of apparent defect density profiles using staircase charge density profile, as proposed by Kimerling and Decock et al.[14, 28] Decock et al. proposed to use this approximation to explain the measurement results of apparent doping density profiles in CIGS solar cells (Eqs. (13), (16) and (17) in ref. [14]), but they did not directly compare the modeling results of this approximate model to their experimental data and to iterative numerical simulations.
The DLS-active defects in CIGS solar cells are complexes. We optimize the following four parameters: (1) the built-in barrier of the CdS/CIGS heterojunction, (2) the shallow doping density , (3) the concentration of DLS-active defects (the transition energy level is ), and (4) the density of the additional acceptor defect (the energy level is ). The optimized parameters are listed in Table A1, together with all the other parameters that are used for the calculation of apparent defect density profiles.
| Materials | ZnO/CdS/CIGS |
| Temperature (K) | 200 |
| Thickness of the CIGS layer (um) | 1.0 |
| Hole effective mass of the CIGS | |
| Built-in barrier of the CdS/CIGS heterojunction (V) | 0.5 |
| (cm) | |
| DLS-active defects | complexes |
| (cm) | |
| (eV) | 0.19 |
| (cm/s) | |
| (eV) | 0.33 |
Table A1. Overview of the material parameters of CIGS solar cells that we use for the calculation of apparent defect density profiles.
We present in Figure A3 the calculated apparent doping density profiles for the CdS/CIGS heterojunction without bias treatment (blue solid line) and after a reverse bias treatment of -2 V (red solid line). The square and triangle symbols represent the numerical simulation and experimental data from Decock et al., respectively; the blue and red colors represent those with and without reverse bias treatment, respectively. The calculated apparent doping density profiles are in satisfactory agreement with the results from Decock et al.

Figure A3. Calculated apparent defect density profiles at the CdS/CIGS heterojunction without bias treatment (blue solid line) and after a reverse bias treatment of -2 V (red solid line). Solid lines are the modeling results using the approximation of staircase charge density profiles. The square and triangle symbols represent the numerical simulation results and experimental data from Decock et al., respectively. Blue and red represent with and without reverse bias treatment, respectively.
For the heterojunction after reverse bias treatment, there is a sharp increase of the apparent doping density (red solid line) when the depletion width reduces below 0.63 um, creating a local maximum of the apparent defect density. The reason is that, in the heterojunction, the length of the distribution of metastable states, that we define as in the main text, is 0.63 um. When the depletion width is larger than 0.63 um, is fixed at 0.63 um and DLS-active defects don’t respond to the capacitance-voltage testing. When the depletion width shrinks below 0.63 um, Decock et al. assume that these DLS-active defects start to respond to the capacitance-voltage testing as regular shallow dopants (the applied equation for the apparent defect density changes from Eq. (17) to Eq. (16) in ref. 14), and as a consequence, the apparent density will jump higher; also, is considered equal to the depletion width and instantaneously changes upon external bias. A more accurate model would consider the dynamics of upon external bias, such as Eq. (25) in this work.
Calculation of the junction resistivity
To calculate the junction resistivity, we consider thermionic emission , diffusion-drift current , and tunneling current . The total current density that flows through the junction is given by:
From thermionic-emission theory we have:43
where is the electron effective mass in the conduction band minimum (CBM), and the Schottky barrier height which has been shown in Figure 2 in the main text. We calculate the diffusion-drift current using the following expression:43
where is the electron mobility and is the conduction band level which has been given by Eq. (20). We calculate the tunneling current density using the Tsu Esaki mechanism:44
where and is the normal component of the electron momentum that is perpendicular to the heterojunction. has been shown in Figure 2 which refers to the energy difference between and the Fermi level . is ratio of the transmitted to the incident current, and using WKB approximation, has been given by Harrison and Stratton:45,46
where is the barrier potential which we compute from the conduction band energy profile , and and are classical turning points what satisfy .
Combining Eqs. (30)-(34), we compute the total current density within the voltage range [-10 mV, 10 mV], and then we are able to calculate the junction resistance (per area) by linearly fitting the total current density versus the voltage.
In the calculation of the resistivity ratio, we assume that the Schottky barrier height is constant in both LRS and HRS. In fact, due to the image-force lowering effect, the resistivity ratio will be larger than the result calculated here. We denote the image-force-induced lowering is . As shown in Figure 3, the interface electric field in the LRS is larger than that in the HRS. Thus, (Eq. (32) in page 148 of ref. 43), and the Schottky barrier height in the LRS is lower than that in the HRS. This will further increase the resistivity ratio compared to the case in which the Schottky barrier is constant in both LRS and HRS.
Accuracy of the approximate expression of
In Eq. (27), we develop an approximate expression of by expanding it linearly with respect to at zero bias. in Eq. (27) is given by the following expression:
Using Eq. (35), we calculate eV/nm.
The exact values of can be computed by combining Eq. (11) and Eq. (20). In Figure A4, we compare the approximate values of versus and which are computed from Eq. (27), with the exact values of which are computed by combining Eq. (11) and Eq. (20). In Figure A4a we fix , and compute and plot versus within the interval [ nm, nm], i.e., [11.1 nm, 21.1 nm]. In Figure A4b we fix nm, and compute and plot versus in the interval of [-1.8 V, 1.8 V]. From Figure A4 we conclude that Eq. (27) approximates fairly well.

Figure A4. Comparison between the approximate values of (represented by red squares) that are computed from Eq. (27) with the exact values of (blue solid lines) that are computed by combining Eq. (11) and Eq. (20). In (a) we fix , and compute and plot versus in the interval of [ nm, nm], i.e., [11.1 nm, 21.1 nm]. In (b) we fix nm, and compute and plot versus in the interval of [-1.8 V, 1.8 V].
References
- [1]J.G. Simmons, R.R. Verderber, and N.F. Mott, “New conduction and reversible memory phenomena in thin insulating films,” Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences 301(1464), 77–102 (1967).DOI
- [2]H.J. Hovel, and J.J. Urgell, “Switching and Memory Characteristics of ZnSe - Ge Heterojunctions,” Journal of Applied Physics 42(12), 5076–5083 (2003).
- [3]A. Odagawa, H. Sato, I.H. Inoue, H. Akoh, M. Kawasaki, Y. Tokura, T. Kanno, and H. Adachi, “Colossal electroresistance of a Pr0.7Ca0.3MnO3 thin film at room temperature,” Phys. Rev. B 70(22), 224403 (2004).DOI
- [4]C.J. Kim, and I.-W. Chen, “Effect of top electrode on resistance switching of (Pr, Ca)MnO3 thin films,” Thin Solid Films 515(4), 2726–2729 (2006).DOI
- [5]C. Pearson, J.H. Ahn, M.F. Mabrook, D.A. Zeze, M.C. Petty, K.T. Kamtekar, C. Wang, M.R. Bryce, P. Dimitrakis, and D. Tsoukalas, “Electronic memory device based on a single-layer fluorene-containing organic thin film,” Applied Physics Letters 91(12), 123506 (2007).DOI
- [6]H. Yin, A. Kumar, J.M. LeBeau, and R. Jaramillo, “Defect-Level Switching for Highly Nonlinear and Hysteretic Electronic Devices,” Phys. Rev. Applied 15(1), 014014 (2021).DOI
- [7]H. Yin, A. Akey, and R. Jaramillo, “Large and persistent photoconductivity due to hole-hole correlation in CdS,” Phys. Rev. Materials 2(8), 084602 (2018).
- [8]D.V. Lang, and R.A. Logan, “Large-Lattice-Relaxation Model for Persistent Photoconductivity in Compound Semiconductors,” Phys. Rev. Lett. 39(10), 635–639 (1977).DOI
- [9]P.M. Mooney, “Deep donor levels (DX centers) in III-V semiconductors,” Journal of Applied Physics 67(3), R1–R26 (1990).DOI
- [10]L. Gordon, J.L. Lyons, A. Janotti, and C.G. Van de Walle, “Hybrid functional calculations of DX centers in AlN and GaN,” Phys. Rev. B 89(8), 085204 (2014).
- [11]M. Igalson, and H.W. Schock, “The metastable changes of the trap spectra of CuInSe2-based photovoltaic devices,” Journal of Applied Physics 80(10), 5765–5769 (1996).DOI
- [12]P. Zabierowski, U. Rau, and M. Igalson, “Classification of metastabilities in the electrical characteristics of ZnO/CdS/Cu(In,Ga)Se2 solar cells,” Thin Solid Films 387(1), 147–150 (2001).DOI
- [13]S. Lany, and A. Zunger, “Light- and bias-induced metastabilities in Cu(In,Ga)Se2 based solar cells caused by the (VSe-VCu) vacancy complex,” Journal of Applied Physics 100(11), 113725 (2006).DOI
- [14]K. Decock, P. Zabierowski, and M. Burgelman, “Modeling metastabilities in chalcopyrite-based thin film solar cells,” Journal of Applied Physics 111(4), 043703 (2012).DOI
- [15]H. Schroeder, V.V. Zhirnov, R.K. Cavin, and R. Waser, “Voltage-time dilemma of pure electronic mechanisms in resistive switching memory cells,” Journal of Applied Physics 107(5), 054517 (2010).DOI
- [16]M. Wang, W. Wang, W.R. Leow, C. Wan, G. Chen, Y. Zeng, J. Yu, Y. Liu, P. Cai, H. Wang, D. Ielmini, and X. Chen, “Enhancing the Matrix Addressing of Flexible Sensory Arrays by a Highly Nonlinear Threshold Switch,” Advanced Materials 30(33), 1802516 (2018).DOI
- [17]M. Cucchi, S. Abreu, G. Ciccone, D. Brunner, and H. Kleemann, “Hands-on reservoir computing: a tutorial for practical implementation,” Neuromorphic Computing and Engineering 2(3), 032002 (2022).DOI
- [18]Z. Wang, S. Joshi, S.E. Savel’ev, H. Jiang, R. Midya, P. Lin, M. Hu, N. Ge, J.P. Strachan, Z. Li, Q. Wu, M. Barnell, G.-L. Li, H.L. Xin, R.S. Williams, Q. Xia, and J.J. Yang, “Memristors with diffusive dynamics as synaptic emulators for neuromorphic computing,” Nature Materials 16(1), 101–108 (2017).DOI
- [19]Y. Zhang, W. He, Y. Wu, K. Huang, Y. Shen, J. Su, Y. Wang, Z. Zhang, X. Ji, G. Li, H. Zhang, S. Song, H. Li, L. Sun, R. Zhao, and L. Shi, “Highly Compact Artificial Memristive Neuron with Low Energy Consumption,” Small 14(51), 1802188 (2018).DOI
- [20]S. Kim, S. Choi, and W. Lu, “Comprehensive Physical Model of Dynamic Resistive Switching in an Oxide Memristor,” ACS Nano 8(3), 2369–2376 (2014).DOI
- [21]S. Dirkmann, M. Ziegler, M. Hansen, H. Kohlstedt, J. Trieschmann, and T. Mussenbrock, “Kinetic simulation of filament growth dynamics in memristive electrochemical metallization devices,” Journal of Applied Physics 118(21), 214501 (2015).DOI
- [22]Y. Li, M. Zhang, S. Long, J. Teng, Q. Liu, H. Lv, E. Miranda, J. Suñé, and M. Liu, “Investigation on the Conductive Filament Growth Dynamics in Resistive Switching Memory via a Universal Monte Carlo Simulator,” Scientific Reports 7(1), 11204 (2017).DOI
- [23]W. Wang, A. Bricalli, M. Laudato, E. Ambrosi, E. Covi, and D. Ielmini, “Physics-based modeling of volatile resistive switching memory (RRAM) for crosspoint selector and neuromorphic computing,” in 2018 IEEE International Electron Devices Meeting (IEDM), (2018), p. 40.3.1–40.3.4.DOI
- [24]W. Sun, B. Gao, M. Chi, Q. Xia, J.J. Yang, H. Qian, and H. Wu, “Understanding memristive switching via in situ characterization and device modeling,” Nature Communications 10(1), 3453 (2019).DOI
- [25]W. Wang, E. Covi, Y. -H. Lin, E. Ambrosi, A. Milozzi, C. Sbandati, M. Farronato, and D. Ielmini, “Switching Dynamics of Ag-Based Filamentary Volatile Resistive Switching Devices—Part II: Mechanism and Modeling,” IEEE Transactions on Electron Devices 68(9), 4342–4349 (2021).DOI
- [26]S.B. Zhang, S.-H. Wei, and A. Zunger, “Intrinsic n-type versus p-type doping asymmetry and the defect physics of ZnO,” Phys. Rev. B 63(7), 075205 (2001).DOI
- [27]S. Lany, and A. Zunger, “Anion vacancies as a source of persistent photoconductivity in II-VI and chalcopyrite semiconductors,” Phys. Rev. B 72(3), 035215 (2005).DOI
- [28]L.C. Kimmerling, “Influence of deep traps on the measurement of free-carrier distributions in semiconductors by junction capacitance techniques,” Journal of Applied Physics 45(4), 1839–1845 (1974).
- [29]C.G. Van de Walle, “Defect analysis and engineering in ZnO,” Physica B: Condensed Matter 308–310, 899–903 (2001).DOI
- [30]J. Dong, Y. Li, Y. Zhou, A. Schwartzman, H. Xu, B. Azhar, J. Bennett, J. Li, and R. Jaramillo, “Giant and Controllable Photoplasticity and Photoelasticity in Compound Semiconductors,” Phys. Rev. Lett. 129(6), 065501 (2022).DOI
- [31]S. Lany, and A. Zunger, “Intrinsic DX Centers in Ternary Chalcopyrite Semiconductors,” Phys. Rev. Lett. 100(1), 016401 (2008).
- [32]G. Chicot, P. Muret, J.-L. Santailler, G. Feuillet, and J. Pernot, “Oxygen vacancy and EC − 1 eV electron trap in ZnO,” Journal of Physics D: Applied Physics 47(46), 465103 (2014).DOI
- [33]T. Grasser, “Stochastic charge trapping in oxides: From random telegraph noise to bias temperature instabilities,” Microelectronics Reliability 52(1), 39–70 (2012).DOI
- [34]Y.-Y. Liu, F. Zheng, X. Jiang, J.-W. Luo, S.-S. Li, and L.-W. Wang, “Ab Initio Investigation of Charge Trapping Across the Crystalline–Si–Amorphous-SiO2 Interface,” Phys. Rev. Appl. 11(4), 044058 (2019).DOI
- [35]A. Alkauskas, Q. Yan, and C.G. Van De Walle, “First-principles theory of nonradiative carrier capture via multiphonon emission,” Phys. Rev. B 90(7), 075202 (2014).DOI