Investigation of atomic layer deposition methods of Al₂O₃ on n-GaN

Due to their better conductance characteristics and ability to maintain high off-state blocking voltages, vertical GaN-based switching transistors are currently receiving increased attention as an alternative to Si- and SiC-based devices.^1–5 GaN-based metal insulator semiconductor field effect transistors, MISFETs, such as vertical trench MISFETs, planar MISFETs, and finFETs, are among the most promising wide bandgap power switching transistor architectures due to their very high ratio between gate width and active area; the channel resistance contributes very little to the overall on-state resistance.^6,7 Unlike Si and SiC, the native oxide layer of GaN may not serve as an efficient gate insulator due to its insufficient electrical properties. Therefore, a foreign gate insulator is obligatory. When considering a foreign gate insulator for reliable GaN-based MISFET switching devices, the following properties need to be kept in mind: (1) secure off-state operation with a wide gate positive voltage span up to 20 V between the off and on-states; (2) threshold voltage stability in safe normally off operation and low forward bias threshold voltage shift; (3) highly conductive on-state channel formed under positive gate bias with high carrier density formed by strong accumulation or inversion; (4) fast switching with low switching losses, i.e., steep subthreshold slope and large on–off ratio; (5) high positive on-state bias robustness and high breakdown strength under off-state blocking conditions. For these reasons, in many proposed GaN-based switching devices, an atomic layer deposited (ALD) amorphous Al₂O₃ layer is used as the gate insulator. Such layers feature a relatively high permittivity, wide bandgap, and a high breakdown electric field.^8–12 In general, two types of ALD methods are used: Thermal Atomic Layer Deposition (ThALD) and Plasma-Enhanced Atomic Layer Deposition (PEALD).¹⁰ In ThALD, the temperature is the main driving force for the layer-forming reaction process,¹³ while plasma-generated reactive species with higher energy are used in PEALD.¹⁴ Schilirò et al.¹⁵ compared the morphological, electrical, and compositional properties of ultra-thin Al₂O₃ insulator layers deposited via ThALD and PEALD on AlGaN/GaN heterostructures. They found that O₂ PEALD offers higher V_ON values and a smoother morphology. They attributed those differences to the better ALD nucleation mechanism provided by the O₂ plasma action. In addition, Qin and Wallace¹⁶ demonstrated using x-ray photoelectron spectroscopy (XPS) that PEALD slightly oxidizes the underlying AlGaN surface during the initial growth steps.

In this work, we identify advantages and drawbacks provided by each method and suggest a novel approach that utilizes the benefits of both ALD process methods: a stacked combination of ThALD and PEALD, where each Al₂O₃ layer is 5 nm thick. We will show that ThALD-based layers are susceptible to an electric field and a low threshold voltage shift, and that PEALD-based layers are more robust when subjected to positive stress bias and feature a higher breakdown field. On the other hand, ThALD-based MISCAPs showcase a lower hysteresis while the PEALD-based MISCAPs show significant hysteresis. Therefore, we will show that combining ThALD and PEALD allows us to utilize both of their advantages while minimizing their disadvantages. In addition to the ALD deposition method, we also investigate the effect of a NH₃ and H₂ plasma treatment of the GaN surface prior to ALD, which is said to reduce the interface state density between GaN and Al₂O₃.^9,17 For GaN-based MIS structures, it has been shown that the highest quality is obtained using a NH₃ plasma pretreatment in combination with PEALD.¹⁸ In this work, we fabricate planar MIS test capacitors (MISCAPs) and characterize their capacitance–voltage (CV) and capacitance–frequency (Cf) characteristics. This is an established method for gate insulator qualification.^19,20

To obtain more insights into the physical–chemical nature of the resulting Al₂O₃–GaN interface, in vacuo XPS measurements are performed after every surface treatment, i.e., pretreatment and ALD precursor step. It is observed that (1) a heat-treated GaN surface is contaminated with carbon and oxygen, which are not removed during the ALD process steps; (2) H₂ plasma pretreatment is beneficial for removing carbon but not oxygen from the pristine GaN surface; (3) NH₃ plasma pretreatment is beneficial for removing carbon and oxygen from the pristine GaN surface; (4) the water precursor step during ThALD results in an incomplete removal of carbon originating from the ligand of the aluminum precursor step; and (5) O₂ plasma during the PEALD method is associated with the complete removal of carbon, but also with the strong oxidation of the GaN surface. In this work, we will link these XPS findings to the MIS capacitor electrical characterizations and demonstrate that the most optimal Al₂O₃ gate insulator is obtained when using a NH₃ plasma pretreatment in combination with initial ThALD followed by PEALD Al₂O₃ growth.

The fabrication process of the metal/Al₂O₃/GaN MISCAPs is described in detail in Ref. 20. For this work, six samples with circular planar MIS capacitors with an area of 0.126 mm² were formed on n-GaN grown by metalorganic vapor phase epitaxy (MOVPE) that consists of 2 μm n-GaN:Si (N_d = 3 × 10¹⁸ cm⁻³) on top of 4 μm unintentionally doped (uid)-GaN. The top metal layer consists of TiW/Au. For each capacitor, approximately 15 nm of oxide was deposited (Table I).

Prior to the ALD process, each sample is heated to 400 °C for 15 min under vacuum for dehydration. Next, the pressure is increased to 0.2 mbar and the sample undergoes a NH₃ or H₂ remote inductively coupled radio frequency plasma pretreatment for 15 min at a plasma power of 200 W. After the plasma pretreatment, the pressure is kept constant by flowing N₂ gas through the reaction chamber, and the sample is cooled to 250 °C. Both ThALD and PEALD processes are performed at this temperature, while the PEALD comprises exposure to O₂ plasma with a plasma power of 200 W. The ALD cycle is composed of trimethylaluminum (TMA) as the aluminum precursor and H₂O as the oxygen source. After each ALD step, the reactor is purged using N₂ gas to remove unreacted H₂O or TMA residues and all reaction products. The ALD process stops when the desired oxide thickness is reached. For the stacked combination, the first layer is 5 nm ThALD, the second layer is 5 nm PEALD, and the third and final layer is 5 nm ThALD. All samples went through postmetallization annealing (PMA) at 350 °C for 45 min.¹⁹ 

In vacuo XPS experiments are performed on GaN-on-sapphire samples that have been grown by means of MOVPE together with the MISCAP samples. A dedicated ALD-XPS setup consists of a home-built pump-type ALD reactor with a base pressure of 10⁻⁷ mbar, which is directly connected to a Theta Probe XPS instrument (Thermo Fisher Scientific Inc.), allowing for direct sample transfer between ALD and XPS without vacuum break. Plasma is generated using a remote inductively coupled radio frequency plasma source at a plasma power of 200 W. H₂ (20% H₂ 99.99% in Ar 99.998%), NH₃ (99.96%), O₂ (99.999%), trimethylaluminium (TMA, 98%), and H₂O are introduced into the reactor at a pressure of approximately 5 × 10⁻³ mbar for a pulse time of 60 s. After each pulse, the reactor is again pumped down to its base pressure. For the plasma pretreatment, multiple plasma pulses are used to achieve a total plasma exposure time of 15 min. All pre-treatments are performed at a sample temperature of 375 °C and the ALD processes steps at 250 °C. The sample is transferred from the ALD reactor to the XPS immediately after the pretreatment and each ThALD and PEALD half-cycle to allow us to observe the difference in chemical composition of the surface after every process step. A new blanket GaN-on-sapphire wafer piece is used to investigate each process. The XPS data are analyzed with CasaXPS (Casa Software Ltd.), and calibration of the spectra is achieved by setting the N-Ga component in the N1s spectrum at 397.7 eV.^21–25 Peak fitting is performed using mixed Gaussian (70%)–Lorentzian (30%) line profiles combined with a Shirley background. The data of all corresponding elements are plotted on the same y-scale for comparison. The error on the relative atomic concentrations obtained from XPS is estimated to be roughly 10% of the reported value.²⁶ 

The MISCAPs are characterized by a forward bias step stress cyclic voltammogram (CV)-sweep method, in which they are first swept to minimal bias V_min = −10 V to deplete as much of the pre-existing interface charge as possible, to V_max which starts from 0 V and which is increased in 1 V steps after each sweep until the measurement reaches V_max = 10 V. Each CV cycle consists of a reverse sweep to depletion and a forward sweep to accumulation. The index “i” represents the sequential measurement number with different maximum stress voltages. This method allows an assessment of the insulator charging effects, identification of potential shifts in the flatband voltage (V_FB), and provides insights into the interface charging phenomena.²⁰ The capacitance was measured at 1 MHz and 0.1 V AC amplitude after a step bias time delay of 100 ms, using a HP 4275A multifrequency LCR meter. Figure 1 shows the CV analysis of the MISCAPs detailed in Table I.

It was not possible to measure the wafers deposited with ThALD up to 10 V. NH₃-ThALD already breaks at 9 V and H₂-ThALD at 4 V for all investigated samples. Thus, we can conclude that ThALD causes a low positive bias breakdown strength (V_BR) of 600 V/μm for the NH₃ plasma pre-treatment and of 270 V/μm for the H₂ plasma pre-treatment. On the other hand, all ThALD samples show considerably smaller hysteresis compared to the PEALD samples, indicating that they contain less voltage-stress-induced interface-trapped charge. The samples undergoing PEALD are measured up to 10 V without difficulty, yet they display a profound hysteresis. Figure 1 shows that the ThALD/PEALD stack method serves as a promising compromise between the advantages and disadvantages of each individual method: the capacitors manufactured with the stacked ALD method are easily measured up to 10 V and they show relatively low hysteresis. Stacks with NH₃ plasma pre-treatments yield a slightly smaller hysteresis than those with a H₂ plasma pre-treatment. Table II summarizes all CV parameters quantitatively: the calculated flatband capacitance (C_FB) (the calculation method described by Tadmor et al.²⁰), the extracted initial flatband voltage (V_FB⁰), and the oxide capacitance (C_ox), which is regarded as the maximal measured capacitance. The PEALD samples, NH₃-PEALD and H₂-PEALD, have the highest V_FB⁰, indicating that they contain more deep oxide trapped charge. The V_FB⁰ of sample H₂-ThALD is relatively close to zero, meaning that if these conditions were applied to an equivalent transistor, they could interfere with its normally off operations.²⁰ All H₂ plasma pre-treated samples tend to have a higher oxide capacitance than the NH₃ plasma pre-treated samples.

In previous work by our group,²⁰ we demonstrated that the deep oxide trapped charges, Q_FB = C_FB × V_FB, are responsible for a V_FB positive shift. With the increase of the positive bias stress, the total induced trapped charge, ΔQ_it, increases. We distinguished between forward and reverse scan induced trapped charge, ΔQ_{it_forward} and ΔQ_{it_reverse} respectively, and showed that for a defined positive bias stress, V_i, ΔQ_{it_forward}(V_i) < ΔQ_{it_reverse}(V_i) and V_{FB_forward}(V_i) < V_{FB_reverse}(V_i), which is the result of added oxide charge. A similar effect, which is responsible for threshold voltage instability in GaN-based MISFETs, is seen in this work.²⁰ 

Figure 2 compares the flatband voltage of each sweep V_FB(i), and the amount of stress-induced trapped charge, ΔN_it, as a function of positive stress bias alongside the applied electrical field E_ox for reverse and forward sweep of the CV measurements. The method for the ΔN_it calculation is explained in detail in Ref. 20. For the NH₃ plasma pre-treatment at forward sweep, it is observed that NH₃-ThALD in comparison to NH₃-PEALD has a lower amount of deep oxide trapped charges, a much lower V_FB(i) shift, and a lower ΔN_it. In contrast, NH₃-PEALD demonstrates a superior forward bias robustness. In the reverse scan, NH₃-PEALD shows a much larger V_FB(i) and a ΔN_it shift compared to NH₃-ThALD. This would be translated to a low positive bias robustness for ThALD and high threshold voltage instability in PEALD. Again, it is observed that the NH₃-stack method displays the advantages of both the NH₃-ThALD and NH₃-PEALD methods. Similar to the NH₃-ThALD layer, it has a higher amount of deep oxide trapped charges, a low V_FB(i) shift, and a low ΔN_it. But unlike the NH₃-ThALD, it owns a good forward bias robustness and E_ox just like the NH₃-PEALD. When comparing the NH₃ and H₂ plasma pre-treatments, some distinct differences are observed. NH₃-ThALD is more robust than H₂-ThALD to forward bias stress and its V_FB⁰ is higher, which indicates a larger amount of deep oxide trapped charges; it has a lower V_FB(i) s, a lower amount of ΔN_it in both sweep directions, and a higher C_FB and C_ox. NH₃-PEALD and H₂-PEALD are both robust to forward bias stress and have similar V_FB⁰. Then again, NH₃-PEALD has lower V_FB(i) s and a lower amount of ΔN_it in both sweep directions and a higher C_FB and C_ox. NH₃-stack and H₂-stack both show comparable properties: forward bias robustness and E_ox, C_FB, C_ox, and V_FB⁰. However, the NH₃-stack has a lower V_FB(i) s and a lower amount of ΔN_it in both sweep directions, which is necessary for stable transistor threshold voltage.

Figure 3 shows the interface traps density of states (D_it) of all samples. The D_it calculation method is described in detail in Ref. 20. D_it estimation is performed using the dynamic conductance method by Nicollian and Goetzberger.²⁷ Such procedure uses frequency modulation from low frequency (1 kHz) to high frequency (5 MHz) to investigate the shallow trap response. This method has been proven as a suitable way to detect both interface trap density and time constant for a wide energy spectrum provided that the leakage current is low enough to not compromise the conductance measurement. In the range between 0.03 and 3 eV, NH₃-ThALD and H₂-ThALD have a similar level of D_it. NH₃-ThALD contains a large concentration of shallow traps, ∼0.01 eV below the conduction band. The NH₃-stack and H₂-stack show a very similar profile with a low D_it between 0.01 and 3.3 eV below the conduction band. When comparing the PEALD samples, NH₃-PEALD and H₂-PEALD, a large difference is observed. H₂-PEALD contains a high D_it in the range between 0.01 and 1 eV. Unfortunately, no direct correlation between the D_it profiles and the electrical properties measured by CV could be found.

To observe the impact on the chemical composition of every surface treatment, i.e., plasma pre-treatment and the first ALD half-cycles, in vacuo XPS measurements are performed after every process step. The pristine n-GaN samples are pre-treated in the ALD reactor, transferred via vacuum in less than 1 min toward the XPS, and measured and transferred back via vacuum to the ALD reactor to undergo the next process step. This method is repeated after every process step and allows us to observe the differences in chemical composition after each process step. The key results can be found in Fig. 4 (selected data after pre-treatment of the n-GaN surface), Fig. 5 (selected data after the first TMA exposure), and Fig. 6 (selected data after the first H₂O or O₂ plasma exposure). The interested reader is referred to Figs. S1–S6 in the supplementary material for a systematic overview of all measured XPS spectra.

Regarding the impact of the GaN pre-treatment prior to ALD, it can be observed in Fig. 4 that both the NH₃ and H₂ plasma pre-treatments successfully remove all spurious carbon contamination from the GaN surface and that a heat treatment only is not sufficient. Only NH₃ plasma is observed to remove a large fraction of the native oxide species that are formed on the GaN surface after air exposure. This is consistent with the results of Yang et al.²⁸ who also observed a significant decrease in carbon and oxygen intensity with XPS after an in situ NH₃ plasma clean. Similar to the results found for H₂ plasma in this work, King et al.²⁹ reported that H₂ plasma could efficiently remove carbon, but not oxygen from AlN and GaN surfaces. In this work, after both NH₃ and H₂ plasma, a new peak is observed in the N1s spectra at 398.7–399.6 eV, which indicates the formation of N–H species at the surface.^21,23,24,30 These species are persistent in the case of the NH₃ plasma pre-treatment as they are not removed from the surface upon TMA exposure in contrast to the H₂ plasma pre-treatment (Fig. 5). It is unclear what the effect of these N-H species might be on the electrical properties, but the abundancy of these species is relatively small with an atomic concentration of only 2%–4%.

After the first TMA exposure, an aluminum peak can be observed at a binding energy of 73.9–75.1 eV (Fig. 5). Simultaneously, an increase in carbon is observed at a binding energy of 283.7–284.3 eV. Both observations are consistent with the adsorption of TMA molecules to the surface. This indicates that we are sensitive to detect the adsorption of a single (sub)monolayer of TMA on the surface.

Next, the TMA-treated surface is exposed to one single H₂O pulse for the ThALD process and to one single O₂ plasma pulse for the PEALD process. Independent of the used pre-treatment, the key results are the same. First, it is observed in Fig. 6 that Al–O bonds are formed on the surface. The observed binding energy of 73.9–75.1 eV agrees with the values reported in the literature for the growth of Al₂O₃ ALD.^15,31 This observation demonstrates how sensitive the in vacuo XPS method is as we can observe the growth of one single Al₂O₃ monolayer on the surface. Second, it is observed that H₂O exposure results in the incomplete removal of carbon from the adsorbed TMA ligands in contrast to O₂ plasma exposure, which is found to remove all carbon from the surface. Similarly, Levrau et al.³² demonstrated by in situ FTIR that H₂O is not reactive enough to remove Si–CH₃ groups that formed after the first TMA pulse from a porous silica surface during ThALD while O₂ plasma during PEALD is. Also, during TiO₂ ALD, Vandenbroucke et al.³³ mentions the incomplete removal of ligands during ThALD vs PEALD. The presence of these carbon species during ThALD might lead to inhibited growth and the incorporation of carbon impurities inside the final film and on the Al₂O₃/GaN interface. In the literature, carbon impurities at the interface are found to be responsible for the presence of interface traps,^34,35 and, therefore, the difference in carbon content between the ThALD and PEALD might have a severe impact on the electrical performance.

Another important difference that can be seen in Fig. 6 concerns the atomic percentage of oxygen, which only slightly increases by approximately 2%–5% after H₂O exposure, whereas after O₂ plasma exposure, the oxygen content increases tremendously by approximately 20%–25%. This increase coincides with the formation of O–Al, O–Ga, and NO–Ga species on the surface detected at binding energies of 530.6–531 eV, 531.2–532.2 eV, and 533–533.4 eV respectively.^{25,31,36–38} The latter binding energy might also indicate the formation of O–H species.³⁹ However, the detection of a component at 399.1–399.5 eV in the N1s spectra supports the identification of Ga–ON species on the surface.^36,40 Possibly, oxygen is substituted onto a nitrogen site during O₂ plasma exposure and acts as a shallow donor interface defect.⁴¹ In anticipation of the discussion of the electrical results, the presence of oxygen defects and creation of less carbon traps might explain why the ThALD/PEALD stack method displays better electrical properties compared to ThALD or PEALD only.

NH₃- or H₂ plasma pre-treatment prior to ALD proves to have a tremendous influence on the MISCAPs electrical properties. Fig. S7 in the supplementary material displays CV measurements for nonplasma pre-treated MISCAPs prior to the ThALD/PEALD stack deposition. It is observed that compared to the NH₃ plasma pre-treated sample, the oxide capacitance is lower, the depletion capacitance is much higher, the reverse sweep hysteresis is significantly larger, and the accumulation slope is less steep. In addition, in Fig. S8 in the supplementary material, it is seen that the shallow D_it density (lower than 0.1 eV below the conduction band) is significantly higher for the nonplasma-treated sample. These phenomena can be understood with the in vacuo XPS results, in which we see that all plasma pre-treatments efficiently remove carbon from the GaN surface. Furthermore, there are several advantages to the NH₃ plasma pre-treatment over the H₂ plasma pre-treatment, in terms of forward bias flatband voltage shift, ΔV_FB(i), ΔN_it, and robustness. This is supported by the in vacuo XPS results, showing that while both tested plasma pre-treatments (NH₃/H₂) efficiently remove carbon contamination from the GaN surface, only NH₃ plasma can remove the native oxide layer of the GaN.

The ΔN_it, V_FB⁰, and V_FB(i) shifts are largely influenced by the different pre-treatments and deposition methods. In contrast, with the exception of H₂-PEALD, no large differences in D_it were identified. We can deduce that D_it does not influence forward bias robustness seen in Fig. 3. Also, there is no definite correlation between D_it and forward bias flatband voltage shift, and the ΔN_it. For example, it is evident from the NH₃ pre-treatment case that the D_it profile for all three deposition methods is relatively equivalent [see Fig. 3(a)].

The MISCAPs’ electrical characterizations presented in Sec. III A indicate that both ALD methods, ThALD and PEALD, have distinct advantages and disadvantages. On one hand, ThALD yields a very low forward bias stress-induced hysteresis and low robustness to positive bias. On the other hand, PEALD yields opposite traits: high hysteresis and high robustness to positive bias. The in situ XPS results in Sec. III B seem to correlate with the large hysteresis displayed by the PEALD samples, by demonstrating a significant increase in oxygen content after the O₂ plasma step. However, in contrast to the water pulse step of ThALD, the O₂ plasma step of PEALD successfully removes all carbon-containing ligands of the TMA precursor step. As such, the O₂ plasma creates one problem, the creation of oxygen defects at the GaN surface, while solving another, the removal of carbon traps. Nevertheless, the in situ XPS results do not provide a definite insight into the low positive bias robustness of the ThALD process. It is hypothesized that in the ThALD/PEALD stack method, the O₂ plasma step reduces the carbon contamination in the ThALD film, while the ThALD film protects the GaN surface from the aggressive oxidation by the O₂ plasma, limiting the formation of oxygen defects, resulting in an improved forward bias robustness.

In this work, we have systematically investigated the effects of plasma pre-treatments of the GaN surface (NH₃/H₂) and ALD deposition methods (ThALD, PEALD, and ThALD/PEALD stack) on the electrical and chemical properties of an Al₂O₃ gate insulator layer for GaN MOSFET devices. The electrical data successfully prove that the newly proposed ThALD/PEALD stack method combines the individual advantages of each deposition method while minimizing the impact of their disadvantages. The in vacuo XPS results of each surface treatment step allow us to explain some of the electrical properties, mainly the native oxide reduction ability of a NH₃ plasma pre-treatment and the dualistic effect of O₂ plasma exposure.

To conclude, the proposed Al₂O₃ ThALD/PEALD stack with NH₃ plasma pre-treatment is suggested as the most suitable gate insulator deposition process for large gate periphery vertical GaN trench MOSFET for fast switching applications.^7,42

See the supplementary material for all in vacuo XPS spectra, measured after every process step; CV measurements of a no plasma vs a NH₃ plasma pre-treated MISCAPs manufactured with the stacked ALD method; and the density of trap states for the no plasma and plasma (NH₃ and H₂) pre-treated MISCAPs manufactured with the stacked ALD method.

This work was funded by the KDT JU (Grant No 101007229). The JU receives support from the European Union's Horizon 2020 research and innovation program and Germany, France, Belgium, Austria, Sweden, Spain, Italy. SENTECH Instruments GmbH is acknowledged for depositing the gate insulator. The authors would like to thank Professor Catherine Dubourdieu from Helmholtz Zentrum Berlin (HZB) for the fruitful discussions. We thank Dr. Jin Li for his assistance with the XPS measurements.

The authors have no conflicts to disclose.

L. Tadmor, S.S.T. Vandenbroucke, and E. Bahat Treidel contributed equally to this work.

Liad Tadmor: Formal analysis (equal); Investigation (equal); Methodology (equal); Visualization (equal); Writing – original draft (equal); Writing – review & editing (equal). Sofie S. T. Vandenbroucke: Formal analysis (equal); Investigation (equal); Methodology (equal); Visualization (equal); Writing – original draft (equal); Writing – review & editing (equal). Eldad Bahat Treidel: Formal analysis (equal); Investigation (equal); Methodology (equal); Visualization (equal); Writing – original draft (equal); Writing – review & editing (equal). Enrico Brusaterra: Formal analysis (equal); Investigation (equal); Methodology (equal); Writing – review & editing (supporting). Paul Plate: Investigation (supporting); Methodology (supporting). Nicole Volkmer: Investigation (supporting); Methodology (supporting). Frank Brunner: Investigation (supporting); Methodology (supporting). Christophe Detavernier: Funding acquisition (supporting); Methodology (supporting); Resources (equal); Writing – review & editing (supporting). Joachim Würfl: Funding acquisition (supporting); Methodology (supporting); Resources (equal). Oliver Hilt: Funding acquisition (supporting); Methodology (supporting); Writing – review & editing (equal).

The data that support the findings of this study are available from the corresponding author upon reasonable request.