Article pubs.acs.org/cm
Mechanisms for Substrate-Enhanced Growth during the Early Stages of Atomic Layer Deposition of Alumina onto Silicon Nitride Surfaces Luca Lamagna,† Claudia Wiemer,*,† Michele Perego,† Sabina Spiga,† Jesús Rodríguez,‡ David Santiago Coll,‡ Maria Elena Grillo,‡,§ Sylwia Klejna,§ and Simon D. Elliott§ †
Laboratorio MDM, IMM-CNR, via C. Olivetti 2, 20864 Agrate Brianza (MB), Italy Centro de Quimica, IVIC, Apartado 21827 Caracas 1020 A, Venezuela § Tyndall National Institute, University College Cork, Lee Maltings, Cork, Ireland ‡
ABSTRACT: The atomic layer deposition (ALD) of aluminum oxide (Al2O3) from trimethylaluminium and water on silicon nitride was studied on as-received and HF-cleaned Si3N4 surfaces. In situ spectroscopic ellipsometry during ALD, X-ray photoelectron spectroscopy, X-ray reflectivity, and time-of-flight secondary ion mass spectrometry were used to elucidate the growth rate, the chemical composition, and the density of Al2O3. The effect of the substrate cleaning and of the growth temperature -varied in the 150−300 °C rangewere analyzed by considering first-principles calculations of the early stages of the growth on both Si3N4 and SiO2 surfaces. Our work evidenced how not only complete ALD cycles but also complementary non-ALD reactions can account for the observed peculiarities related to the enhanced or inhibited growth rates on the Si3N4 surfaces as a function of temperature. KEYWORDS: atomic layer deposition, in situ ellipsometry, first principles calculations, trimethylaluminium, Al2O3
I. INTRODUCTION The increasing demand for more compact memory storage devices from a variety of electronic applications motivates the development of nonvolatile flash memories based on the charge trapping mechanism of silicon nitride as in the SONOS device structure.1,2 In the SONOS structure, the charge is trapped within a layer of amorphous silicon nitride (Si3N4), inserted between two silicon oxide layers (SiO2), which act as tunnel dielectric and blocking layer to prevent charge injection from the control gate. A high dielectric constant (high-k) dielectric, such as alumina (Al2O3), may alternatively be used as blocking oxide; this yields a good programming efficiency while allowing a thicker tunnel oxide and avoiding the retention issue.3,4 In this way, a higher field on the tunnel oxide will be obtained for the same applied voltage. Still, compatibility of the materials at the interface can be crucial to the desired mechanical and electrical properties. One of the key issues for the integration of materials into the memory stack is good electrical quality of the interface between Al2O3 and Si3N4. Hence, a fundamental understanding of the deposition process is required in order to tailor the quality of the Al2O3 film and its interfacial properties. To this end, the early stages of the deposition of Al2O3 on silicon nitride and oxide surfaces have been studied in the present work by both experimental and theoretical techniques. Atomic layer deposition (ALD) is the technique of choice for conformal deposition of oxide films at the nanometre-scale. For the deposition of Al2O3 using trimethylaluminum (TMA, AlMe3, Me=CH3) and water (H2O),5 the net growth reaction is 2AlMe3(g) + 3H2O(g) → Al2O3(s) + 6CH 4(g) (1) © 2012 American Chemical Society
This is one of the more reliable ALD processes, with an optimum growth rate of 1.1−0.9 Å/cycle at 200−300 °C.6,7 Our aim is to characterize and understand the ALD growth process on a Si/SiO2/Si3N4 substrate, layer-by-layer at the atomic level, with the goal of controlling interface formation e.g., via use of appropriate surface cleaning or deposition temperature. We consider here Al2O3 on Si3N4 substrates, but the approach may be applicable to other systems. We specifically seek to understand how substrate preparation affects the subsequent reaction with precursors (TMA and H2O) and what the consequences are for interfacial characteristics. Many open questions remain about the reaction of ALD precursors with technologically important substrates during the early stages of ALD growth. In general, the initial conditions of the substrate are found to have a strong influence on the growth process. In the so-called initial transition regime, growth has been commonly identified as linear growth, inhibited growth or substrate-enhanced growth.5 It should be noted that the transition regime is sometimes restricted to just a few cycles, making detection of the interface reactions very difficult unless in situ monitoring of adequate sensitivity is employed. Here, we propose three possible chemical scenarios that may be at play during the growth from cycles of pulsed TMA and H2O. (i) The preparation of the substrate may affect the rate of the standard, self-limiting ALD reaction of TMA, primarily via the coverage of hydroxyl groups on the substrate (the reactive sites in eq 2 below). (ii) TMA is a reactive molecule and can be Received: November 11, 2011 Published: February 8, 2012 1080
dx.doi.org/10.1021/cm203362d | Chem. Mater. 2012, 24, 1080−1090
Chemistry of Materials
Article
effect of the substrate on the growth reactions and growth rates of ideal ALD. Obtaining a higher growth rate than the prediction indicates that the reaction of TMA with the surface is not selflimiting and that a different mechanism is contributing to the growth of the oxide. Alternative growth mechanisms are also investigated with DFT.
expected to undergo other (non-ALD) reactions depending on the chemical nature of the substrate, either contributing extra product or etching the substrate away. (iii) Steady-state ALD of Al2O3 onto Al2O3. In the simplest case, substrate effects are limited to the earliest few ALD cycles, until a single monolayer of interfacial layer (IL) is formed, which is then followed by steady-state growth. However, the situation may be complicated if substrate and product mix to form a more extended IL with its own distinct growth chemistry, as is often the case when using an aggressive oxidizing agent such as ozone (O3).8−10 Experimentally, the growth behavior is generally investigated via the rate of deposition per cycle. A linear growth already from the early stages of the deposition is the ideal ALD case, and it can be observed for metal oxides and metal nitrides grown on SiO2.11 On the other hand, inhibited initial growth is generally associated with island growth and is therefore characterized by an incubation period that is necessary to activate the surface reactions or to let the islands grow into a complete layer (e.g., TiN, Ru).11,12 In this paper, ultrathin Al2O3 films are deposited using TMA and H2O on Si3N4 surfaces at different temperatures. The growth rates of the first few Al2O3 monolayers are determined by in situ spectroscopic ellipsometry (SE). Moreover, X-ray photoemission spectroscopy (XPS), time-of-flight secondary ion mass spectroscopy (ToF SIMS), and X-ray reflectivity (XRR) are used ex situ to assess the chemical composition, thickness, and electron density of the substrates and films. We report the rate of alumina growth per ALD cycle in two ways: the thickness growth rate (TGR), expressed in Å·cycle−1, and the molar growth rate (MGR), expressed in Al2O3·nm−2·cycle−1. The MGR can be converted to the TGR by dividing by the density of Al2O3 units in the film (Al2O3·nm−3) and multiplying by ten. Experimental13 and theoretical14 evidences indicate that the following reaction mechanism underlies the ALD of Al2O3 from TMA and H2O. In both precursor pulses, the first step is chemisorption of the precursor onto unsaturated sites on the surface of the growing film: H2O onto Lewis acidic surface Al (surf-Al) and TMA onto Lewis basic surf-O or surf-OH. The adsorbed precursor may dissociate via further unsaturated sites (protons from H2O transferring to surf-O or Me ligands transferring to surf-Al).14 Combination of the surface-bound moieties is the reaction by which ligands are eliminated from the surface
II. METHODS 2.1. Experimental Method. 2.1.1. ALD and in Situ SE. Al2O3 thin films were grown in a Savannah 200 ALD reactor (Cambridge Nanotech, Inc.) using TMA and H2O as Al and oxygen source, respectively. The growth temperature (Tg) was varied between 150 and 300 °C. The ALD growths were performed on a three layers stack (i.e., Si/SiO2/Si3N4). The Si3N4 layer, nominally 60 Å thick, was grown on Si(100) substrates covered with a 45 Å nominally thick SiO2 layer. Two different surfaces were explored: as-grown bare (“as-received”) and HF-cleaned Si3N4 (“cleaned”) substrate surfaces. The Al2O3 film thickness target was set to ∼20 Å. The thickness analyses of the substrates before the ALD growth were performed employing SE and XRR. From the combination of the two techniques, the thickness (±2 Å) of the SiO2 bottom layer was fixed at 45 Å, while a 61 Å thick Si3N4 layer was evidenced. Chemical cleaning of the Si3N4 surface was performed, aimed at removing any possible oxide on top of the nitride surface. The samples were dipped in a HF aqueous solution (HF:H2O = 1:50) for 30 s, rinsed in deionized water, and dried using a N2 flow. In this case, for the HF cleaned stack, SE and XRR analyses revealed a slightly reduced Si3N4 thickness of 57 Å. This finding suggests that most likely a successful etching of the oxidized top layer has occurred. To specifically address the very early stages of film deposition, the film growth rate evolution was carefully investigated by in situ SE after the H2O pulse and purge of each of the first 20 ALD cycles. SE measurements were performed using an M2000-F ellipsometer (J. A. Woollam Co., Inc.) sending the incident light beam through two quartz windows installed on the ALD reactor lid. The beam angle of incidence with respect to the substrate normal was 70°; the photon energy range employed was between 1 and 5 eV. The experimental Ψ and Δ spectra, collected pulse by pulse during the ALD depositions, were analyzed using the EASE and WVASE sofware (J. A. Woollam Co., Inc.).15,16 The ellipsometry data can be expressed in terms of the pseudodielectric function.17 Prior to the film deposition, the pseudodielectric function of the as-received and HFcleaned substrates was extracted by means of a Point-to-Point interpolation of the spectra also including the temperature dependence of the optical constants of the silicon substrate. This enabled us to model the Si/SiO2/Si3N4 stack exactly in the experimental conditions at the beginning of each ALD growth. Al2O3 dielectric function was estimated using a Cauchy dispersion relation18 (oxide is transparent in the 1.0−5.0 eV range) on thicker samples (i.e., 20 nm thick films) and was then applied to model the thin films. Therefore, SE analysis was performed assuming fixed Al2O3 optical constants. The only free parameter of the SE fit was the Al2O3 film thickness. From in situ SE measurements both the film thickness and TGR evolution are monitored and represented as a function of the number of ALD cycles. TGR was calculated taking into account the neat thickness variation measured by in situ SE after each ALD cycle. If the chemical composition of the surface changes, the TGR monitored by in situ SE should be expected to vary along ALD deposition, and thus cycle by cycle, until a steady-state regime is achieved. 2.1.2. Chemical and Structural Characterization. XPS measurements were performed on a PHI 5600 instrument using a conventional X-ray source (monochromatic Al Kα = 1486.6 eV) and a concentric hemispherical analyzer with a nominal energy resolution of 0.5 eV. The spectra have been collected at a takeoff angle of 45°. ToF-SIMS depth profiles were acquired on a dual-beam IONTOF IV system. Sputtering was accomplished by Cs+ at 0.5 keV over a 200 × 200 μm2 area, while the analysis has been performed in negative polarity by using Ga+ primary ion beam operating at 25 keV and
surf‐Al(Me)n + surf‐OH → surf‐Al(Me)n ‐ 1 + surf‐O + CH 4(g)
(2)
Assuming facile diffusion of ligands/protons, this elimination reaction will proceed, freeing up sites for further adsorption. In the TMA pulse, all residual protons from the previous H2O pulse will eventually be eliminated from the surface and a saturating coverage of Me ligands will be obtained. At this point, there are no unsaturated sites for TMA adsorption. This self-limiting chemistry in the TMA pulse ensures an upper limit on the amount of Al deposited per cycle, which is characteristic of ideal ALD. From the point of view of the substrate, we can identify requirements for the ideal ALD mechanism to occur: Lewis acidic/basic sites for chemisorption and the capacity to produce surface protons that are sufficiently Brønsted acidic for elimination. We therefore use density functional theory (DFT) to compute the surface acidity and hydroxyl coverage of nitride and oxide surfaces, both bare and hydroxylated, in order to estimate the 1081
dx.doi.org/10.1021/cm203362d | Chem. Mater. 2012, 24, 1080−1090
Chemistry of Materials
Article
rastering over a 50 × 50 μm2 area. ToF−SIMS data were normalized on the value of 30Si signal in the silicon bulk in order to remove variations of signal intensity due to fluctuations of the Ga current. XRR analyses were performed using a system and method already described elsewhere.10 The XRR spectra were simulated by using a three layer model: the Al2O3 layer on top of the Si3N4/SiO2/Si substrate. The electron density, surface roughness, and thickness of the Al2O3 layer, as well as the thickness and surface roughness of the Si3N4, were set as free parameters during the fitting procedure, while the other parameters were fixed at the values measured on the two different (as-received and cleaned) substrates. The obtained layer thicknesses were found in all cases to be in very good agreement with the thicknesses extracted by means of SE, thus validating the employed optical model. 2.2. Theoretical Method. In this section the theoretical method used to develop models for the as-received and cleaned substrates is presented. We consider hydroxylated surfaces of the oxidized and cleaned substrates i.e., surfaces of SiO2 and Si3N4, respectively, with a degree of hydroxylation that may depend on temperature. All bulk and surface structures were calculated using the generalized gradient approximation to DFT of Perdew, Burke, and Ernzerhof (PBE)19 and the all-electron projector-augmented wave (PAW) method20 as implemented in the VASP package.21 Other DFT codes were used for specific analysis, as noted below. 2.2.1. Models of Bulk Crystals. The total energy versus volume of bulk α-SiO2 in the crystalline α-quartz trigonal structure D36 was optimized with respect to the crystallographic lattice parameters a and c/a and to the internal coordinates of the 9 atoms (6 O and 3 Si) per unit cell. A grid of (3 × 3 × 3) special points for Brillouin zone integrations and a kinetic energy cutoff of 400 eV was adopted in the calculations. The optimized lattice parameters a of 4.980 Å and c/a of 1.09 thereby obtained are consistent with the experimental values22 of 4.913 Å and 1.10, respectively. The hexagonal primitive cell of β-Si3N4 with space group P63m was also optimized with respect to the crystallographic lattice parameters and to the four inner parameters of the coordinates of the 14 atoms (6 Si and 8 N) per unit cell. A detailed description of the computational setup and account of the resulting bulk structure has been published elsewhere.23 A model of α-Al2O3 was optimized with a plane-wave basis up to a cutoff of 500 eV and 48 k-points in the Brillouin zone (4 × 4 × 2 Monkhorst-Pack grid). Geometry optimization of both ion and cell coordinates was performed to a convergence of less than 10−3 eV. The resulting equilibrium lattice parameters agreed well with experiment: a = 4.808 Å, c/a = 2.731 (experimental values: a = 4.759 Å, c/a = 2.730).24 2.2.2. Models of Substrate Surfaces. The natural cleavage plane (0001) of β-Si3N4, consisting of nearly planar layers of stoichiometric Si3N4, was adopted to model the cleaned substrate. The bulk truncated β-Si3N4(0001) surface exhibits unsaturated three-coordinated Si and two-coordinated N sites. These undercoordinated Si and N atoms readily react with atmospheric water to form the groups Si−OH and N−H. Therefore, a hydroxylated β-Si3N4(0001) surface model was developed as representative for the cleaned substrate. The surfaces were modeled by periodic slabs comprising seven layers along the [0001] axis (8.8 Å thick), separated by 12 Å of vacuum, or 15 Å of vacuum when TMA was included as an adsorbate. This slab, which consists of three bulk repeat-units along the c-direction, provided a reasonably well converged surface energy as a function of thickness, converging the atomic forces to an accuracy of less than 0.2 meV·Å−2 between successive optimization cycles. For the sake of calculating the surface energy for the asymmetric slab,25 the three topmost layers defined a surface region and the three bottom layers of the slab a bulk region. The atom coordinates within the surface region were optimized employing five special points for Brillouin zone integrations and a cutoff of 400 eV for the plane-wave basis set. The (111) plane of silica in the α-quartz structure was selected for the as-received surface motivated by the small surface lattice mismatch (
