Introduction

The transmission electron microscope (TEM) has been used for many years to investigate the morphology and atomic structure of microelectronic device structures, while their analytical properties—electronic structure and chemical composition and bonding—have been the subject of surface-sensitive, medium-resolution techniques: ellipsometry for oxide thickness and dielectric determinations, deep-level transient spectroscopy and capacitance­voltage measurements for defect electronic structure, X-ray diffraction for crystal strain, and secondary ion mass spectroscopy (SIMS) for composition. A device structure imaged using a TEM could be identified by inference with such analytical properties.

Today we face the problem that devices are too small for reliable, routine diagnosis using surface-sensitive, low-resolution probes. For instance, it is very difficult to investigate the composition of a DRAM capacitor trench sidewall by means of SIMS. The uncertainties in probed area, secondary ion yield for non-normal probe incidence, and the buried nature of the interface all limit the lateral resolution to about 50 nm and the depth resolution to about 5 nm [3]. The many new techniques derived from the scanning tunneling microscope (STM) satisfy some of those needs, particularly for dimensional metrology, but suffer from their surface-oriented nature when confronted with a morphologically complicated shape [4].

In this paper, we present a brief overview of the use in our laboratory of the scanning transmission electron microscope (STEM) to fulfill some of those needs. It is fundamentally analytical in nature in that it is designed not only to determine atomic structure, but also to obtain information on the composition and atomic bonding environment of a specimen. Scanning transmission electron microscopy, like SIMS and scanning electron microscopy, is based on the use of an electron probe. It shares with the TEM a high-resolution capability to locate atom positions; it shares with SIMS an ability to distinguish the kind of atoms that are present and how they are bonded; and it shares with the scanning electron microscope (SEM) a simplicity of interpretation grounded in its scanning nature. As we see below, under some circumstances it can also provide atomic-resolution electronic structure information similar to that obtained from large areas by optical absorption. It is also fundamentally a bulk-sensitive probe—it can, for example, be used to obtain information within a localized but subsurface region that controls the operation of a microelectronic device.

Annular dark-field imaging in the STEM

Figure 1

Figure 2 depicts the ADF imaging method. Shown there is a schematic illustration of the electron-scattering geometry in the STEM. Electrons are scanned in a raster across the specimen; they travel through the specimen and are caught by various detectors which feed video displays for viewing. The STEM is therefore a combination of a high-resolution TEM instrument with a low-resolution SEM. In ADF imaging, we rely on the fact that an electron which passes near an atom in a specimen will suffer a large electrostatic deflection. If we place a detector well away from the undeflected beam direction, we can obtain a signal using only the electrons that have been deflected through large angles and which have therefore passed close to an atom column. If the probe is smaller than the distance between the atom columns, about 0.3 nm for the [0 1 1] orientation of Si, this detection method yields a large change between the on-column and off-column positions of the probe. The detected intensity is roughly proportional to the number of atoms in the column times the square of their atomic number [9]. Thus, we have a signal that gives us information about the locations of the atoms, their concentration, and their atomic numbers.

Figure 2

This imaging technique was pioneered in the 1970s by Crewe, who was the first to show atomic resolution and single-atom sensitivity with atomic number contrast using a scanned electron probe [6]. More recently, the technique has received extensive investigation for imaging interface and defect structures in crystals [7, 8]. Because the technique is sensitive to the atomic number (Z) of the atom, ADF imaging is sometimes referred to as “Z-contrast.” Therefore, brighter areas are those composed of more atoms or atoms having a higher atomic number. In Figure 1, the Si matrix and Si₃N₄ dielectric are more dense and therefore brighter than the SiO₂ dielectric. Because an ADF STEM signal averages over many scattering angles, there is less contrast arising from diffraction than is normally observed in a TEM. In the figure, the Si matrix surface is above; the poly-Si-filled trench capacitor extends several microns below into the substrate, and is bounded by an SiO₂/Si₃N₄/SiO₂ dielectric insulating layer along the trench sidewalls. There is a poly-Si strap connecting the top of the capacitor with the Si matrix region that would become the drain end of a FET in a completed device. The contrast in this image is primarily due to material density, with the most dense areas causing the most scattering into the high-angle detector.

Figure 3 is a composite of higher-magnification ADF images of the trench dielectric. Intensities are roughly proportional to atomic density, so that the Si matrix and poly-Si capacitor appear with the same brightness. Variations reflect grain boundaries and orientation differences in the poly-Si. SiO₂ appears darker than Si. Within the 15-nm dielectric, we can image three regions—associated with the presence of SiO₂, Si₃N₄, and SiO₂. Bright-field (BF) imaging in this area would show no variation in contrast within the dielectric. There appears to be a 2­3-nm-thick region of brighter contrast following the edge of the Si matrix. Identification of this requires higher magnification and atomic-resolution analytical information, to be discussed next.

Figure 3

Atomic column imaging

Figure 4

In this image, each spot corresponds to a double column, with the column separated by 0.14 nm in the [1 1 0] projection. Since the STEM probe diameter in this experiment was 0.2 nm, the columns are not resolved, appearing instead to be single spots elongated in the [1 0 0] direction (left­right) in the image. Returning to the DRAM structure, Figure 5 shows a further magnification of the small bracketed area in the inset of Figure 3. This image clearly shows elongated spots ending at the Si/SiO₂ interface at the trench sidewalls. We can now see that the narrow region of bright contrast in the lower-magnification images does indeed coincide with the Si lattice. The crystal structure appears relatively unmodified except for the ADF intensity variation. Further understanding of this result requires an analytical analysis of bonding or composition, as discussed below. At this point, however, it can certainly be inferred that atomic-resolution images of technologically interesting structures are readily obtainable by means of an STEM, with the apparent limit being imposed by instrumental probe size limitations rather than specimen preparation.

Figure 5

Analytical analysis using electron energy loss spectroscopy

Figure 6 shows an illustrative EELS spectrum for Si in the 0­120-eV energy loss range. For a sample 50­100 nm thick, most of the incident electrons do not lose any energy and therefore end up in the no-loss peak centered at 0 eV energy loss. The second large peak near 17 eV results from electrons that lose energy to one collective electronic excitation, or plasmon. This excitation involves all of the valence electrons in the solid and so is characteristic of electron density, but is less informative about the bonding or electronic structure. Near 100 eV for Si, an absorption edge is evident for electronic transitions from the 2p core levels to the CB. This feature is only about 1% as strong as the plasmon peak. As summarized in the magnified insert, there is fine structure within 2 to 3 eV of the core loss onset energy, giving information about the CB in the solid. Finally, in the 1­3-eV region, near the no-loss peak, there is a direct interband absorption edge present which is caused by energy loss via direct excitation of single electrons across the bandgap. This is only about 0.01% as large as the no-loss peak and is therefore very difficult to observe [11]. High-energy-resolution optical spectroscopy has been obtained in this energy range by cathodoluminescence (CL) using micron-sized electron beams [12], and energy-integrated CL has been obtained in the STEM using small beams [13]. However, EELS is the only STEM-based signal that promises to obtain both energy and spatial resolution to allow investigation of a semiconductor bandgap, and so it is currently the subject of experimental and instrumental investigation [14, 15].

Figure 6

Therefore, core-to-CB absorption has enjoyed the most widespread use as an analytical signal because it has been relatively easy to obtain and is usually quite specific to composition. In Figure 7 several Si 2p_3/2 CB spectra are shown for different Si bonding environments. These shapes can be used to identify the Si compounds with the lateral spatial resolution limited by the probe size even when analyzing single atom columns [16]. We can use this to understand the trench sidewall image in Figure 5, where EELS from three areas are displayed in the figure insert. They show clear signatures for the bulk Si and SiO₂ in regions a few atomic layers away from the interface. Within the narrow transition region, a composite shape is observed, one consisting partly of Si and partly of SiO₂, in spite of the fact that a clear Si atomic lattice is present. It seems likely, therefore, that this interface is not smooth in the [1 1 0] imaging projection. This might happen if oxygen is driven into the interface during cleaning, or possibly if the cross-sectioned region did not precisely cut the curved trench sidewall perpendicular to its surface. This kind of imaging and spectroscopic behavior has also been observed by others at the Si/SiO₂ interface in thin gate dielectrics, where interface roughness has been cited as a possible cause [17].

Figure 7

At a somewhat more detailed scale, the elemental Si absorption edge itself can be interpreted for CB information. In Figure 8 two Si 2p_3/2 absorption edges are shown. These edges have been extensively processed by 1) subtraction of a slowly varying background, 2) deconvolution of an experimental instrumental resolution function, and 3) removal of structure derived from the spin-orbit splitting of the spin 3/2­1/2 core states. The data are then compared with model spectra derived from trial s,d-symmetry-projected LDOS curves. These curves are generated analytically, assuming that each symmetry-allowed band edge may be expanded in a parabolic fashion about the band minimum. The band minimum and carrier effective mass are then considered to be fitting parameters for comparison with the measured data.

Figure 8

Figure 9 shows a typical band structure for Si [18]. There are three main symmetry-allowed CB final states: 1) the s-like minimum at ₁, dispersing in the [1 0 0] direction; 2) the s-like minimum in the [1 1 1] direction L₁; and 3) the d-like saddle point at L₃ in the [1 1 1] direction. Data for the Si core absorption are plotted over the diagram to show how the features of the spectra correspond to features of the bands. We can create a trial LDOS for the comparison to the data by summing LDOS along the allowed bands. If a band is particularly flat in energy, an inflection point is generated in the LDOS. Finally, experimental effects are included, such as distortion by the core exciton, core lifetime broadening, and instrumental resolution to produce the solid curve fit to the experimental data in Figure 8. This spectral processing is discussed in detail in [19].

Figure 9

As is clear from Figure 8, this procedure gives excellent reproducibility. The two sets of experimental results in that figure were from the Si substrate and Si capping layer in a Ge₃₀Si₇₀ quantum well structure [20]. The positions of ₁, L₁, and L₃ are clearly reproduced, with better than 20-meV accuracies. In the relaxed Ge₃₀Si₇₀ alloy, these major CB points have shifted systematically. We have followed the evolution of these values throughout the relaxed Ge_xSi_1­x alloy series and find, as shown in Figure 10, that the CB follows very closely the predicted behavior as a function of Ge content [21]. In fact, we find on completion of this exercise that the Si 2p core level is constant in energy in the relaxed alloy, allowing the use of the absorption edge position to track the band offset in this system [22].

Figure 10

With some minor assumptions regarding strain, we can even use these results to obtain the band offset in a strained quantum well [23]. This is illustrated by Figure 11, in which we have obtained the heterojunction CB offset as a function of position across a 9-nm strained Si well. In the figure, we show an ADF image of the well which shows the atom columns and identifies the position of the well by atomic number (Z-contrast). A line scan of the recorded intensity along the dotted line is shown in overlay. The position of the CB is obtained from EELS spectra and is also recorded on the image. This structure contains a deliberately introduced grading of the composition on the substrate side of the well, readily visible in the ADF line scan on the left. The band offset obtained clearly follows this composition variation on the left as well. At the upper interface (to the right), where quantum well conduction is desired, the composition is made as discontinuous as possible. A surprising result is that the EELS CB offset variation is sharper than the composition profile at the upper interface. This could be the result of “over-strain” at the interface, a condition that has been suggested previously to be present in Si­Ge multilayers, but which has not been conclusively identified [24, 25].

Figure 11

How might this very fine structure be exploited in the case of the DRAM structure of Figure 1? This particular structure was fabricated to test the electrical integrity of the strap connection between the drain of a DRAM access transistor and a poly-Si region of a DRAM capacitor. It was found that under some circumstances, the resistance of the strap connection appeared to be 100 times bigger than normal. However, TEM inspection of the structure showed no obvious reason for the behavior. In Figure 12(a), the Si/strap interface from the upper portion of Figure 1 is shown at high magnification in ADF imaging. The Si substrate is on the left; the poly-Si strap is on the right. Si 2p absorption spectra are shown at the lower right. It is clear from the image that there is a 1-nm-thick region of low density at the interface. The 2p CB absorption results show that this region is largely composed of Si, although a minor amount of O may be present. The Si edge shape is distinctly bulklike on either side of the interface, but within the interface is less sharp, with the loss of fine structure attributable to the bulk bands. This behavior has been observed in fine Si particles and in other Si objects in confined situations [26, 27]. It appears therefore that the poly-Si growth interface, although primarily Si in composition, is disordered on a 1-nm scale, leading to a lower density and to damping of the CB structure by electron confinement effects. Inspection of the ADF image 1­2 nm deep into the Si substrate provides a possible explanation for this behavior. Si stacking fault (SF) structures are visible (circled), suggesting that the reactive ion etch (RIE) cleaning, used to prepare the Si surface for growth of the poly-Si strap, caused subsurface damage that led to the development of surface stress that inhibited growth.

Figure 12

Figure 12(b) shows an example of an analysis of a low-resistance strap interface. This interface has an almost uniform density with no discernible interface layer. In fact, the interface is rough at the 0.5-nm level, indicating intimate contact between the Si substrate and the poly-Si strap. The EELS spectra show bulklike fine structure throughout. No subsurface defects are visible. It is intriguing to note that the interface does show some blurring of the CB fine structure, suggesting that the interfacial quality is not yet as good as it could be, and that some improvement in electrical characteristics may therefore be possible with further work on this interface.

Conclusions and future directions

Acknowledgments

References