Lateral gallium-nitride (GaN) field-effect transistors are currently employed as switches for efficient power conversion. While lateral GaN high electron mobility transistors (HEMTs) are already commercially available, vertical counterparts are still in a relatively early development phase, and intensive research is undergoing to harness their advantages over the lateral devices in high power applications^{[1, 2]}. These advantages stem from the fact that in vertical transistors the drift region develops through the device depth instead of length. This design allows an increase in the maximum blocking voltage without extending the device laterally and allows for the peak electric field to be far from the device surface, enhancing both stability and reliability^{[2, 3]}. Consequently, vertical devices operating safely in the kV range can be designed with a relatively small chip area and are naturally immune from common drawbacks that vex lateral devices operation (e.g., current-collapse, dynamic R_on, etc.)^[3−7]. Furthermore, reliable and unclamped inductive switching in converter circuits employing vertical GaN devices can be achieved thanks to their avalanche capability^[8−10]. However, vertical GaN devices are yet to be fully optimized and suffer from stability and reliability issues of their own^[11], particularly regarding the gate oxide and its interface with GaN^{[2, 11−13]}.

In this work, we investigated the influence of gate oxide optimization on the transfer characteristics hysteresis (H) and subthreshold-slope (SS) of two pseudo-vertical GaN-on-Si trench MOSFETs (TMOS’s), labelled device A and B, with different gate stack deposition technique. Experimental characterization was carried out to quantify the hysteresis and its dependence on the maximum gate−source voltage applied, whereas TCAD simulations were adopted to gain insight into the observed H and SS reduction obtained with device B with respect to device A. Our analysis allows to entirely attribute the presence of the hysteresis to electron trapping into border traps, i.e., to trap states located in the gate oxide at a tunneling range from the oxide/semiconductor interface, and the observed H improvement from device A to device B to a 4 × reduction in the border-trap density. SS is instead mainly governed by interface traps and its improvement can be ascribed to a 2 × decrease in their density. Moreover, the mobility improvement between the two devices can be attributed to increased SiO₂/GaN interface quality, which reduces the amount of interface traps.

Devices under test (DUT’s) in this work are pseudo-vertical GaN trench MOSFETs (TMOS’s) on 150 mm GaN-on-Si wafers. The MOS gate stack was based on an LPCVD SiO₂ gate dielectric with polysilicon gate electrode which has previously been shown to be suitable for GaN planar MOSFETs^{[14, 15]}. The schematic 2-D device structure is shown in Fig. 1. The epitaxial GaN stack from bottom to top consists of a strain-relief buffer layer and a n⁺ drain layer (1 µm) to achieve ohmic contact with the drain metal. The drain contact is deposited on the surface laterally with respect to the device active mesa region which consists of a n⁻ drift region (1 µm), a p body channel region (0.5 µm), and a n⁺ source layer (0.3 µm). The nominal magnesium concentration in the p body layer is 10¹⁹ cm⁻³. It should be noted that the thickness of n⁻ drift region is too low to block considerable voltages. Indeed, its thickness was chosen such that its resistance contribution is minor compared to the channel resistance. The gate trench was formed by chlorine plasma etching followed by wet treatment in TMAH to reduce crystal damage induced by the dry etch process and selectively prepare the m-plane of the GaN crystal.

The deposition of the 70-nm silicon dioxide (SiO₂) single-layer gate dielectric differed in the two devices (A and B) considered in this work. For device A, gate oxide was deposited by low-pressure chemical vapor deposition at 785 °C with a post-deposition anneal at 800 °C, whereas for device B, SiO₂ deposition was carried out at 880 °C with a post-deposition anneal at 900 °C. It was previously found that such an annealing treatment improves the field effect mobility in the channel of lateral channel MOSFETs^[15].

Fig. 1(b) shows a TEM image of the gate trench for device B. It can be seen that fabrication resulted in straight trenched sidewalls with a smooth interface between semiconductor and dielectric perpendicular to the c-planes, as also shown by the HRTEM image at 820 Kx magnification in Fig. 2. Both device A and B were characterized with Keysight 5260 source measurement units (SMU's) by performing a back-and-forth gate-to-source voltage (V_GS) sweep in DC regime while keeping the drain-to-source voltage (V_DS) fixed at 1 V and recording the drain current (I_D). A set of I_D−V_GS transfer characteristics were acquired by increasing the maximum applied voltage (V_GS,max) from 5 to 40 V; the minimum applied V_GS was always –5 V. The sweep rate was 1 V/s, and the bias step was 100 mV. Note that 40 V is the maximum allowed bias to avoid excessive gate leakage, while typically maximum bias during device operation is ≤20 V.

To restore the state of the device and avoid trapped charge accumulation, after each sweep the DUT’s were exposed to 365 nm (3.4 eV) ultraviolet (UV) illumination for a few minutes. Experimental data are collected in Fig. 3, showing the set of I_D−V_GS curves for device A (a) and B (b) for different V_GS,max. Saturation current (I_D,SS) at V_DS = 5 V and V_GS = 20 V is 0.12 mA/mm and 0.2 A/mm for device A and B, respectively. The main factors limiting I_D,SS are the mobility degradation due to electron conduction near the semiconductor/oxide interface and the Fermi level pinning due to charged acceptor interface and border traps.

The almost overlapping threshold voltage (V_T) of the forward I_D−V_GS curves for both devices indicates that the 3.4 eV UV light is effective in restoring the state of the device. On the other hand, at room temperature it was found that 100-s recovery phase at V_GS = 0 V after 100-s stress at V_GS = 40 V led to a residual ≈2 V positive V_T shift, i.e., full recovery was not achieved, see Fig. 4.

For both devices, we observed a V_T shift between the forward and backward sweep, indicating that slow traps get filled with electrons during the forward sweep and are not thermally emitted during the backward sweep. The V_T difference between the backward and forward sweep is commonly referred to as hysteresis. Note that V_T is taken at V_GS for which I_D = 1 μA/mm. Because the amount of trapped charge increases with increasing V_GS,max, hysteresis is also proportional to V_GS,max^[16]. As can be noted by comparing Figs. 3(a) and 3(b), device B hysteresis is remarkedly lower than that of device A. This behavior is further investigated by means of numerical simulations in Section 3.

In general, both electron trapping and detrapping are dynamic processes that determine a positive/negative V_T drift, respectively, and as such these V_T drifts are proportional to the duration of bias application. Non-zero hysteresis can thus be interpreted as the net result of electron trapping followed by an electron release that is not fully completed in the time window of the applied V_GS range.

Fig. 4 better clarifies this point, showing the V_T drift during stress/recovery experiments carried out for several time decades (from 10 μs to 100 s) at different stress voltages (V_GS,STR) and at a fixed 0 V recovery bias (V_GS,REC) on device B. During stress and recovery cycles, stress (recovery) bias was periodically turned off to allow the measurement of the I_D−V_GS curve to probe V_T, by sweeping V_GS from 0 to 30 V in 10 μs with V_DS = 1 V. In this case, due to insufficient resolution of the experimental setup at low currents, V_T was taken at I_D = 1 mA/mm. The non-saturating behavior of V_T shift with both stress and recovery time in Fig. 4 is in agreement with previous studies of PBTI on GaN-based planar devices^[17−19].

Looking, for instance, at the case V_GS,STR = 40 V in Fig. 4, the net V_T drift (taken as the difference between the value at the beginning of the stress cycle and that at the end of the recovery one) is about 2 V, which correlates nicely with hysteresis at V_GS,max = 40 V as shown in Fig. 3 (and Fig. 7). This result suggests that non-zero hysteresis can be interpreted as the net result of the dynamic processes of electron trapping and incomplete recovery (from trapping) in the time window of the applied V_GS sweep.

Furthermore, V_T drift during stress at V_GS,STR = 30 V was characterized at different temperatures to better understand the physical mechanism at the origin of hysteresis. Fig. 5 shows the results of this characterization on device A, and it can be observed that V_T drift during stress is not influenced by temperature, supporting the hypothesis of tunneling as the underlying mechanism. A more detailed discussion of the V_T shift during stress and recovery experiments is provided in Ref. [20].

Two-dimensional numerical device simulations were performed with the Synopsys TCAD tool SDevice^[21] aimed at obtaining insights on the improvement in the stability of I_D−V_GS curves of the two devices discussed in Section 2. Particularly, we investigated the impact of border and interface traps on hysteresis and subthreshold slope reduction. Charging and discharging of interface traps are accounted for by means of a fully dynamic SRH model (no quasi-static approximation), that is self-consistently coupled to the drift-diffusion equations. Additionally, border traps are placed into SiO₂ gate oxide region and are allowed to exchange electrons with the GaN surface through the nonlocal “tunneling into traps” model included in the simulator^[21]. A constant mobility model, which is only dependent on temperature, was adopted for simplicity. Hence, technological development in terms of mobility improvement is reproduced in the simulations by setting the channel mobility value (μ_ch) in the body layer (i.e., where the inversion channel forms) for the two devices to the respective field-effect mobility extracted from the measurements.

The energy distributions and concentration of interface (IT) and border traps (BT), as well as channel mobility, were firstly calibrated to obtain a best fitting with the experimentally measured hysteresis on device A. Box-like distributions for trap distributions were adopted for a more straightforward parametrization. The trap parameters are summarized in Table 1.

The box-like distribution has the following form^[21]:

1{DT,EM−0.5ESEEM+0.5ES0,elsewhere,

where D_T is either D_BT or D_IT (see Table 1). E_M and E_S are defined in Table 1. We assume that the donor interface traps have a large energy distribution across the upper portion of GaN energy gap and that the acceptor interface traps are narrowly distributed near the conduction band edge (E_C), which corresponds to assume that the charge neutrality level^[23] is located 0.1 eV below E_C. Interface and border traps distribution employed in the simulations are shown in Fig. 6. The interface trap distribution was modeled starting from the distribution experimentally validated for other material systems (e.g., Si−SiO₂ interface^[24], Al₂O₃−InGaAs interface^[16]) and conventionally assumed for oxide/semiconductor interfaces^[22]. Border trap distribution parameters were set in order to reproduce experimentally measured hysteresis.

The channel mobility (μ_ch) was set to 3 cm²/(V∙s), corresponding to the field-effect mobility extracted from the measurements. The results of the simulations of device A in terms of I_D−V_GS curves for different V_GS,max (V_DS = 1 V) are shown in Fig. 7(a). Good agreement is obtained between experiments and simulations, compare Fig. 7(a) and Fig. 3(a). Hysteresis is physically induced by channel electron tunneling into the border traps during the upward V_GS sweep and only partial electron release during the downward one, resulting in a right-shifted threshold voltage.

By properly reducing the IT and BT concentrations, as well as increasing channel mobility, without adjusting any other parameter, we calibrated the simulation results on device B as well, as shown by the I_D−V_GS curves in Fig. 7(b). The trap parameters are again reported in Table 1. μ_ch in this case was set to 17 cm²/(V∙s), corresponding to the field-effect mobility extracted from the measurements.

The physical mechanism of hysteresis can be better understood with the aid of Fig. 8(a), showing the simulated band diagram across the channel (i.e., plotted along the lateral dimension perpendicular to the vertical current flow) for different V_GS (see legend). Additionally, Fig. 8(b) shows the concentration of trapped charge in the oxide (n_BT) and free electron density (n) in the channel corresponding to the same biases. As V_GS increases, so does n in the channel as the Fermi level moves above the conduction band edge (i.e., inversion occurs). The raise in Fermi level position (with respect to E_C) increases the probability for electron tunneling from the channel to the traps in the gate oxide (border traps). This consequently increases n_BT in the same region, as shown in Fig. 8(b).

Fig. 9 shows the same quantities as Fig. 8 but taken on the forward and backward curves at the same bias of V_GS = 5 V. The larger n_BT on the backward curve compared to the forward one causes a smaller n in the channel, see Fig. 9(b), leading to higher V_T (and thus, hysteresis). This is consistent with the band diagrams shown in Fig. 9(a), as the band bending corresponding to V_GS = 5 V on the backward curve is less than that on the forward one, meaning that inversion is weaker in the former than in the latter case, thus explaining the reduced n.

The agreement between the simulations and experimental data in terms of hysteresis is further confirmed by Fig. 10, that shows how the calibrated setups are able to capture the linear dependence of hysteresis on V_GS,max for both TMOS devices. Remarkably, the hysteresis reduction observed between the two devices of ≈75% at V_GS,max = 40 V is captured by the simulations when reducing the BT density from 4 × 10¹⁹ eV^–1∙cm^–3 to 1 × 10¹⁹ eV^–1∙cm^–3. As stated earlier, typical gate bias during actual device operation is ⩽20 V so that H in practice is <1 V for device B.

IT concentration has little role on the hysteresis, due to the short emission time constants and fast charge (during upward V_GS sweep) and discharge (backward V_GS sweep)^[16]; however, the IT density reduction favorably impacts SS, besides correlating with the mobility improvement as a result of the interface quality improvement^[25].

By extending the BT and IT density range, it is possible to extrapolate the corresponding H and SS values, particularly to estimate the amount of trap density reduction required to minimize both H and SS. Fig. 11 shows the results of this analysis for the hysteresis vs BT density curve (with all other parameters of the setup calibrated on the experimental data of device B). This analysis reveals a strong dependence of hysteresis on BT density in the 5 × 10¹⁸ − 5 × 10¹⁹ eV^–1∙cm^–3 range: the higher the trap density, the higher the concentration of trapped electrons that cause the V_T shift between forward and backward sweeps. Further reducing BT density below 10¹⁹ eV^–1∙cm^–3 leads to negligible hysteresis.

Fig. 12 illustrates the dependence of SS on the interface trap density. As can be seen, the ≈30% SS reduction obtained by device B compared to device A is explained by TCAD simulations as a result of D_IT reduction.

We presented electrical characterization data and simulation results on two different pseudo-vertical GaN-on-Si trench MOSFETs. The higher temperature gate oxide deposition technique used in device B compared to device A led to a reduced transfer curve hysteresis and subthreshold slope, as well as increased channel mobility. Calibrated simulation decks on both devices allowed us to investigate in detail the cause for hysteresis, which can be attributed to oxide traps near the interface with the semiconductor (border traps). The subthreshold slope is instead mainly governed by the density of the interface traps. The improvements in hysteresis and subthreshold slope achieved by device B can therefore be attributed to a consistent decrease in the density of border and interface traps, respectively. The sensitivity analysis carried out by means of TCAD simulations in terms of hysteresis (subthreshold slope) vs border trap (interface trap) density, allows quantifying the additional trap density reduction required to minimize both figures of merit.