Transconductance Overshoot, a New Trap-Related Effect in AlGaN/GaN HEMTs

DC characteristics of AlGaN/GaN HEMTs with different thickness values of the undoped GaN channel layer were compared. An abnormal transconductance (<inline-formula> <tex-math notation="LaTeX">${g}_{m}{)}$ </tex-math></inline-formula> overshoot accompanied by a negative threshold voltage (<inline-formula> <tex-math notation="LaTeX">${V}_{\text {TH}}{)}$ </tex-math></inline-formula> shift was observed during <inline-formula> <tex-math notation="LaTeX">${I}_{\text {DS}}$ </tex-math></inline-formula>–<inline-formula> <tex-math notation="LaTeX">${V}_{\text {GS}}$ </tex-math></inline-formula> sweep in devices with thinner GaN layer. At the same time, a non-monotonic increase in gate current was observed. In OFF-state, electron trapping occurs in the undoped GaN layer or at the GaN/AlN interface, leading to a positive <inline-formula> <tex-math notation="LaTeX">${V}_{\text {TH}}$ </tex-math></inline-formula> shift. When the device is turning on at a sufficiently high <inline-formula> <tex-math notation="LaTeX">${V}_{\text {DS}}$ </tex-math></inline-formula>, electron de-trapping occurs due to trap impact-ionization; consequently, <inline-formula> <tex-math notation="LaTeX">${V}_{\text {TH}}$ </tex-math></inline-formula> and therefore <inline-formula> <tex-math notation="LaTeX">${I}_{\text {D}}$ </tex-math></inline-formula> suddenly recovers, leading to the <inline-formula> <tex-math notation="LaTeX">${g}_{m}$ </tex-math></inline-formula> overshoot effect. These effects are attributed to electron trap impact-ionization and consequent modulation of the device’s electric field.

electronics, modern telecommunication systems as 5G [1], [2], [3]. Recent requests for higher frequency operation of these devices have pushed the gate length (L g ) scaling down to a technological limit. As L g becomes smaller and the device aspect ratio decreases, short-channel effects (SCE) may become more relevant [4], [5], [6]. Control of SCE requires scaling of vertical dimensions; reduction of substrate conductivity can be achieved by compensating the GaN buffer with Fe and/or C. However, this may give rise to parasitic effects, such as increased gate leakage current [7], [8], [9], [10], [11], decrease of electron mobility [12], trapping phenomena [13], [14], [15], and V TH instability [16], [17]. An alternative way to control leakage current and SCE is to use a not-intentionallydoped GaN channel/buffer with largely reduced thickness, so that the AlN nucleation layer acts as a backbarrier.
In this work, devices under test were fabricated using undoped GaN channel layers of different thicknesses (thin, medium, and thick); interaction between hot electron effects and electron traps lead to a g m overshoot effect which was enhanced in devices having a thin GaN channel. This transconductance overshoot, previously unexplained [18] is therefore another sign of the presence of deep levels in GaN HEMTs, as current collapse [19], [20] or kink effects [21], [22]. A physics-based explanation of this effect is proposed and discussed, which is relevant for the characterization of deep levels effects in scaled GaN HEMTs.
The device under test and experimental details are described in Section II, while results of dc electrical characterization are reported in Section III: a transconductance (g m ) overshoot was observed during transfer I D -V GS measurements at high V DS . Section IV describes the results of threshold voltage (V TH ) transient spectroscopy. Discussion and conclusions are presented in Section V.

II. DEVICE AND EXPERIMENTAL DETAILS
The RF GaN HEMT tested in this work are based on a buffer-free AlGaN/GaN on SiC epitaxy [23], [24], [25] and were processed using a production-level 0.15 µm gate length (L g ) technology. In this study, the thickness of the undoped GaN channel layer was the only variable among the samples analyzed. The samples adopted three different thickness values    [25], [26] ( Fig. 1), which are identified as thin, medium, and thick for ease of reference. The devices under test are two-finger devices with a width (W G ) of 2 × 40 µm.

III. DC CHARACTERIZATION
The output I D -V DS , transfer I D V G , g m -V GS characteristics, and I -V curves of the gate-source diode of devices having different GaN channel thicknesses are shown in Fig. 2(a)-(d).
Results show that the thin devices have the best subthreshold swing (SS) and the smallest drain-induced barrier lowering (DIBL), but slightly reduced drain current (I DSS ).
The thick devices show the largest drain-source leakage current (I DS,leak ), and the highest I DSS , as summarized in Table I. In the buffer-free, ultrathin GaN layer design, the AlN nucleation layer, with a high bandgap, acts as a backbarrier and effectively controls drain-to-source leakage, thus improving short-channel effects. Better electron confinement results in very good values of SS, DIBL, and I DS,Leak of the thin devices. However, the thinner GaN layer implies a higher vertical electric field for the same bias conditions, thus enhancing electron trapping and R ON increase.
All devices are affected by a small kink effect in the output I D -V DS characteristics, Fig. 2(a). Abnormal g m overshoot effects are observed in the medium and thin devices; the g m overshoot amplitude is maximum in the thin devices [see Fig. 2 To check the influence of reverse bias in OFF-state on the overshoot, the g m -V GS characteristics were measured in the saturation region (V DS = 10 V), increasing the absolute value of the starting gate voltage value (V GS,Start ) in pinch-off conditions, (Fig. 3). Measurements have been taken using a 20 ms integration time. When V GS,Start was decreased in steps from −3 to −7 V, the g m overshoot became more and more apparent and was accompanied by a gradual positive V TH shift in the thin and medium devices [see Fig. 3 The presence of hysteresis effects in the g m -V GS characteristics was studied by taking the measurements twice, i.e., first by increasing (V GS ), i.e., from −7 to 1 V ("Go" measurement) and then with decreasing V GS from 1 to −7 V ("Back"), at V DS = 10 V, as shown in Fig. 4(a) for a representative thin channel device; I G -V GS characteristics were simultaneously measured. During the "Go" measurement, once V GS becomes larger than V TH ( ∼ =−2.5 V), electrons populate the channel, drain current starts to flow, and g m increases steeply to the overshoot value. At the same bias conditions, a bell-shaped bump appears in the I G -V GS characteristics, corresponding to a sudden increase of gate current, Fig. 4(b).
In pinch-off conditions, a negative gate current (I GS ) is measured, consisting of electrons injected from the gate into the AlGaN barrier. As V GS increases, |I G | becomes lower, since the reverse bias between the gate and source/drain decreases, leading to lower electric field values. However, at V GS ≈ −2.32 V, |I G | starts to increase again and creates a bell-shaped bump in the I G -V GS characteristics, which is typical of impact-ionization effects [27], [28]. Finally, around V GS = −0.5 V, |I G | resumes the initial decreasing trend.
During the "Back" sweep, no g m overshoot is observed, and V TH recovers (i.e., it shifts toward more negative values), Fig. 4(a). In the "Back" sweep, I G shows no bump, and its absolute value is larger than the values measured during the "Go" phase [see Fig. 4 Conventionally, the non-monotonic increase of |I G | is attributed to the hot-electrons-induced generation of holes via band-to-band impact-ionization (i.i.), and subsequent collection at the gate. As V GS is increased beyond V TH , 2DEG density increases, more hot electrons are generated and can impact-ionize, thus increasing |I G |. For higher V GS values, due to electric field decrease and enhanced phonon-and electronelectron scattering, electron energy and impact-ionization are reduced, leading to the observed non-monotonic, bell-shape behavior of |I G |.
In the devices under test, however, direct band-to-band i.i. seems unlikely: 1) the bell-shaped |I G | increase is observed even at V DS = 4 V, which is too low for band-to-band i.i. to occur in a wide bandgap material like GaN [29] and 2) I G versus V GS characteristics show hysteresis effects which suggest that traps may be involved; in fact, experimental data can be explained by trap assisted impact ionization, consisting in the detrapping of electrons due to the interaction (impactionization) between high energetic hot electrons and filled deep levels, without hole generation. The energy required for this process is just the difference between the conduction band edge and the deep level energy, much lower than the energy gap. As in the case of band-to-band impact-ionization, trap-related impact-ionization requires hot electrons, but it can occur at much lower values of energy (or electric field). When, at increasing V GS , the channel is opened, channel hot electrons impact-ionize traps, and the negative charge stored in the access region between the gate and drain is removed, leading to a sudden increase of the electric field [29], and a consequent jump of the gate leakage current. At the same time, a negative threshold voltage shift occurs, leading to the g m overshoot due to the corresponding sudden increase in drain current. As V GS is further increased, the decrease of V GD and of the electric field induces a decrease in the gate leakage current. The bell-shaped, non-monotonic behavior of I G is therefore due to a modulation of the electric field consequent to trap impact-ionization. In its turn, the increase of the electric field induces an increase in gate leakage current. Since the latter dominates I G its thermal coefficient is positive, Fig. 5(b), with an activation energy of around 0.87 eV. Gate current (I G ) as a function of temperature [see Fig. 5(b)] shows that the bell-shaped behavior is present at all temperatures, from 25 • C to 95 • C.

IV. THRESHOLD VOLTAGE TRANSIENTS
V TH transients were measured using a two-phase trap filling/recovery experiment. During the filling phase (duration 100 s), the device was biased in pinch-off at V GS = −6 V, V DS = 0 V. During the recovery phase, the device was biased either in OFF-state (V GS,B = −4 V) or in the semi-ON state (V GS,B = −2 V). Recovery experiments were repeated at increasing V DS from 1 to 15 V, in 1 V steps. At various intervals during both phases, filling or recovery bias was turned off for 4 µs and the I D V GS characteristics were measured. The "dynamic" V TH value was extrapolated; its values are reported in Fig. 6 for a representative thin device.
During the filling stress, a positive V TH shift is observed, with a logarithm dependence on the stress time. This behavior was previously detected in transistors based on silicon [30] [31], GaN [32], [33] and SiC [34], and is usually described by an "inhibition model" [30], [32] which assumes that when an electron is trapped, a Coulombic potential is generated, and this inhibits charge trapping in neighboring defects, thus decreasing the trapping rate. The long filling time explains why V TH positive shift and g m overshoot are not observed for short integration times (20 µs).
When the recovery phase is evaluated in OFF-state with V DS lower than 4 V, negligible recovery of V TH (<10 mV) occurs [see Fig. 6(b)]; for higher V DS , further trapping occurs and an additional positive V TH shift is observed [35]. On the contrary, when the device is biased in a semi-ON state (V GS = −2 V) with V DS ≥ 3 V, significant recovery occurs, i.e., device current is required to promote detrapping. On the other hand, for V DS ≥ 7 V fast (100 µs) hot-electron trapping effects compete with de-trapping, and recovery is incomplete [see Fig. 6(c)].
The shape of the V TH transients can be explained as follows: When, after an OFF-state stress phase, the device recovery is studied in the semi-ON state, hot electrons are generated, inducing two competing mechanisms: 1) hot electrons may be trapped on the device surface or in the buffer, thus inducing a positive threshold voltage shift and 2) hot electrons impact-ionize negatively charged traps, leading to negative threshold voltage shift (Fig. 7). The former mechanism is very fast (10-100 µs), the second is slower (1-10 ms), leading to a non-monotonic behavior of V TH transients as a function of time, Fig. 6(c). At V DS higher than 8 V, the first mechanism prevails, and V TH does not recover anymore. Fig. 8 shows V TH semi-on recovery transients as a function of temperature in a representative thin device, after 100 s filling at (−6 V, 0 V). During recovery, the device was biased at V GS,B = −2 V, V DS,B = 6 V, in semi-on. The Arrhenius plot [see Fig. 8(b)] indicates that recovery is very weakly thermally activated, with a very low activation energy of 0.05 eV, possibly resulting from trapping and detrapping mechanisms having opposite dependence on temperature. The identification of the relevant deep levels is difficult, as the detrapping mechanism is due to trap impact-ionization and not to temperature-enhanced electron release. This also contributes to making the activation energy of the recovery process extremely low.

V. DISCUSSION AND CONCLUSION
In this work, buffer-free AlGaN/GaN HEMTs with ultrathin GaN layers of different thicknesses were studied. It is shown that by reducing GaN layer thickness, one can effectively improve charge confinement, reducing SS and DIBL. However, trapping effects in devices having a thin GaN layer thickness led to g m overshoot effects, and bell-shaped |I G | increase. These effects can be explained by the interaction of hot electrons, deep levels, and trapped charge.
In OFF-state, electron trapping occurs in the undoped GaN layer or at the GaN/AlN interface, leading to positive V TH shift, as shown in Fig. 9. When increasing |V GS | in OFF-state conditions, trapping is enhanced and further V TH shift occurs, Fig. 3. When the GaN channel layer is thicker, the distance between the 2DEG channel and the defective region at the GaN/AlN interface is increased [23], [29], [36], thus reducing trapping effects within the GaN layer and at its interface with AlN.
When the device is turning on at a sufficiently high V DS , electron de-trapping occurs, due to the impact-ionization of filled traps by hot electrons. V TH (and therefore I D ) suddenly Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply. recovers, leading to the g m overshoot effect during g m -V GS measurements, as shown in Fig. 3.
The de-trapping mechanism is identified by analyzing its correlation with I G : an indirect mechanism involving the impact-ionization of traps can indeed explain all the experimental results.
In GaN HEMTs, negative trapped charge under the gate and in the gate-drain access regions controls the electric field profile. Once electrons are removed by trap i.i., the electric field increases, leading to gate leakage current increase. Trap impact-ionization follows the same non-monotonic trend of band-to-band impact ionization, causing a bell-shape in the I G -V GS curve (Fig. 5). However, once traps are depleted of electrons, they remain empty during the subsequent part of the measurement, as filling time is longer than measurement time (Fig. 6). This explains the hysteresis during "Go" and "Back" I G -V GS measurements (Fig. 3): during the "Go" measurement with V GS from −7 to 0 V, the negative charge is removed as the channel is opened, through traps i.i. As a consequence, during the "Back" measurement (V GS from 0 to −7 V), traps are empty, without trapped negative charge in the gate-drain region, and the electric field becomes higher, thus leading to a higher gate reverse leakage current I G , as shown in Fig. 5.
Trap i.i. is confirmed by several experimental observations: de-trapping is not due to the electric field, as it does not take place in OFF-state, Fig. 6(b); the electron energy required to impact-ionize traps is much lower than that needed for the band-to-band generation of electron-hole pairs, so that trap i.i. and the consequent transconductance overshoot can occur at V DS values as low as 4 V. Finally, semi-ON-state de-trapping is not thermally activated (Fig. 8), possibly due to the compensation of two counteracting mechanisms, i.e., the decrease of electron energy at increasing T , and thermallyactivated de-trapping.
In conclusion, in buffer-free AlGaN/GaN HEMTs adopting a thin undoped GaN channel layer, trapping in OFF-state followed by impact-ionization of traps in semi-ON state explains the anomalous g m overshoot, as well as the observed V TH shift and the non-monotonic, bell-shaped, increase of |I G |.
Devices adopting a thick GaN layer are free from those parasitic effects, showing that optimization of GaN thickness and epitaxial layers leads to excellent performances of these HEMTs in terms of gate length scalability and electrical reliability.