Introduction

Achieving this level of control requires a detailed understanding of the fundamental processes which take place during processing. Research laboratories throughout the world have therefore been investigating reaction mechanisms and the relationship between structure and properties at all levels of integrated circuit fabrication, and as our knowledge becomes deeper, our ability to improve device performance increases. As an example of this, consider the relationship between the quality of an ultrathin gate oxide layer and the initial silicon surface on which it is grown. Studies have shown that interface steps do not move during oxidation [1]; it has therefore become clear that the initial silicon surface must be extremely flat if a uniform oxide with a smooth interface is to be created. In the field of Al- and Cu-based chip metallization, study of the dynamics of voids has led to designs of the lines and vias which mitigate the effect of electromigration [2]. For novel devices such as single-electron transistors or solid-state lasers [3], an understanding of the growth mechanism of nanosize islands, or “quantum dots,” allows these islands to be formed controllably and devices to be fabricated.

Many analytical techniques are available for studying growth and processing during the fabrication of electronic devices.¹ In particular, transmission electron microscopy (TEM), especially in combination with advanced specimen preparation methods [4], is an indispensable technique for examining the layered structures which make up semiconductor devices. TEM has been used successfully, for example, to identify phases in polycrystalline films, to measure film texture, and in the atomic scale imaging of relevant interfaces [5]. However, most TEM investigations have relied on postprocessing examination of device structures. A full understanding of processing must also include the transient effects which occur during heating, exposure to gases, or deposition, and these effects may be time-consuming or impossible to capture by conventional ex situ TEM analysis. We have therefore developed a transmission electron microscope in which many of the processing steps important in device fabrication can be reproduced. By observing growth and phase transitions in situ we can understand their mechanisms and model relevant processes.

In this paper we illustrate the importance of this type of in situ analysis by discussing two experiments. In the first example we describe the growth of nanosize semiconductor islands, or “quantum dots,” by self-assembly during strained-layer epitaxy, and show how in situ growth of islands in the TEM allows us to observe the phenomenon of self-assembly in real time. By observing the evolution of island sizes and shapes, we were able to determine how to optimize processing conditions to achieve the narrow size distribution required for quantum dot devices. Furthermore, observations of island nucleation suggested ways to pattern the islands in particular regions of the specimen. Together these results indicated that self-assembly is a viable technique for forming quantum dots for novel devices. In the second example, we describe studies of a phase transition in the Ti­Si system. The transition from the C49 phase of TiSi₂, which has a high resistivity, to the C54 phase, having a much lower resistivity, must occur consistently during processing of low-resistance contacts. Real-time observations of silicide phase formation in Ti films deposited in situ and ex situ allowed us to measure the growth kinetics of the desirable C54 phase and to show that the propagation of the phase boundary is rate-limited by oxygen and other impurities at grain boundaries in the precursor C49 phase. These results suggested changes in processing conditions which might encourage the phase transition and thereby improve contact properties. We conclude with a discussion of a different class of growth process—electrochemical deposition of materials from a liquid solution—and show how in situ TEM may allow the dynamics of liquid/solid interfaces to be studied.

Design of a transmission electron microscope for the study of growth processes

Figure 1

Silicon substrate specimens were used in the present study. The specimens consisted of 2 × 4-mm rectangles cleaved from 100-µm-thick Si(001) or Si(111) wafers which had been polished on both surfaces. They were thinned to electron transparency with an HF/HNO₃/CH₃COOH mixture. Final thinning was carried out by etching in the microscope polepiece using a low pressure of oxygen (10^­6 Torr) at about 900°C until a desired thickness was achieved [1]. The specimens were then flash-heated to above 1250°C to form a clean surface on which growth could take place. Dynamic processes were recorded in real time using an image intensifier² linked to a video system, or on photographic plates which were subsequently digitized for computer analysis.

The growth and patterning of quantum dots of Ge on Si

One very promising route to the fabrication of nanosize volumes of material is to make use of the phenomenon of island self-assembly during strained-layer epitaxial growth [8]. To illustrate how this occurs, consider the case of Ge growing epitaxially on Si (Figure 2). The Ge lattice spacing is about 4% larger than that of Si, so significant strain energy is stored in the Ge lattice as it grows. As Ge is deposited beyond about three monolayers (ML), the growing surface becomes nonplanar and islands form, relieving strain by allowing the Ge lattice to expand at the island peaks. Depending on the growth conditions, the islands may be up to 100 nm in diameter and can have a very uniform size distribution.

Figure 2

The spontaneous formation of nanosized islands during strained-layer growth, if it can be sufficiently well controlled, provides a means to fabricate quantum dots which bypasses the limitations of lithography. However, the self-assembly process has its own limitations: In particular, the placement of individual islands is not defined at the start of growth, and island sizes may show large variations. Thus, the critical questions in using self-assembly for quantum dot fabrication are as follows: Can we control the size (and shape) of the islands formed, and can they be grown in specific locations? Extensive study of island self-assembly in Ge/Si, as well as in other systems (InP/GaInP, InGaAs/GaAs, and CdSe/ZnSe) in which the phenomenon occurs [9], have led to some highly successful recipes for growing quantum dots with desirable properties. However, the increasing importance of these structures makes it essential to control the size distribution and placement more precisely, and this has stimulated further detailed studies.

Postgrowth examination of island populations has revealed several intriguing aspects of the self-assembly process. One particularly important result is that in many cases islands can be grown with a very narrow size distribution [10]. Although this is very convenient for device applications, the reason for the narrowness of the distribution is not well understood. The distribution is generally considered too narrow to be due to Ostwald ripening [11], and several other growth mechanisms have been proposed. Some models are based on thermodynamic arguments, postulating the existence of a minimum energy size to which islands will tend to grow [12]. Other models suggest that kinetic effects slow down the growth of larger islands, allowing the smaller ones to catch up. Possible reasons for this may be the slowing of diffusion to an island due to the strain field around it [13] or the reduced probability of nucleating a step on the surface of a larger island, again due to strain [14].

A second interesting result, so far observed only in the Ge/Si(001) system, is the recent observation of a bimodal island size distribution, with the two peaks corresponding to two different island shapes [15]. Smaller islands (less than about 60 nm in diameter) have been found to be square-based pyramids, while larger ones adopted steeper “dome” shapes with additional facets. The origin, evolution, and distribution of the pyramids and domes are important for the development of quantum dot devices and are the subject of much debate at present.

We therefore examined the process of island self-assembly in Ge on Si in real time using the in situ TEM system. Ge was grown by CVD on Si(001) at a specimen temperature of 650°C using digermane gas (Ge₂H₆) at a pressure between 10^­8 and 10^­6 Torr, resulting in a growth rate of 0.1­10 monolayers of Ge per minute. A weak-beam imaging condition was chosen which showed the strain field around the islands. This allowed island positions and sizes to be resolved, although the images did not show the shapes of individual islands.

Representative results from the experiments are shown in Figures 3 and 4. The in situ observations revealed a rich evolutionary process. Figure 3 indicates how the bimodal size distribution develops in a large population of islands. Shortly after islands become visible, a broad bimodal distribution develops, with two peaks corresponding (presumably) to the pyramid- and dome-shaped islands. At later times, the smaller (pyramid) peak weakens and almost disappears. While a single postgrowth observation made, for example, after 50 seconds [as in Figure 3(b)] might suggest two stable populations, the complete series shows that the pyramids are transient, with no particular size preferred, and vanish at later times. Note that the domes show their narrowest size distribution just as the pyramids are disappearing [Figure 3(c)].

Figure 3

To explain these results, we have developed a simple model [16] based on the existence of the two known island shapes. During growth, islands can exchange Ge atoms, since at the relatively high growth temperature of 650°C the equilibrium concentration of Ge adatoms on the surface is large (~10% [17]), and Ge atoms can easily detach from an island, diffuse along the surface between the islands, and reattach to another island. Calculations show that this process occurs so readily that the flux of detaching and attaching Ge adatoms is more important in controlling the development of an island than the flux of Ge arriving from the gas phase. If one island, because of its shape or size, has a lower energy per atom than another island, Ge atoms will preferentially detach from higher-energy islands and attach to the lower-energy island. Islands can thus shrink or grow depending on their energy per atom in comparison to the other islands.

If all of the islands were the same shape, we would expect that the larger islands would always have a lower energy per atom than the smaller ones by virtue of their lower surface energy/volume ratio. The larger islands would grow at the expense of the smaller ones, leading to the well-known classical phenomenon of Ostwald ripening [11]. However, if two different island shapes are allowable, the process becomes much more complex and interesting. For the Ge pyramid and dome shapes described above, simple calculations of the energy per atom show that the pyramid shape is preferred for smaller islands below a critical volume V_c, while the dome shape is favored at volumes greater than V_c. The islands formed during the early stages of Ge deposition are all small and therefore pyramid-shaped and undergo Ostwald ripening, with the larger islands growing at the expense of smaller ones. As the successful pyramids grow, the first few to reach the volume V_c make the transition to the dome shape. Their energy per atom rapidly decreases, making them very attractive sites for adatoms and increasing their growth rate still further at the expense of the smaller islands. The result of this process is shown in Figure 4(a), while in Figure 4(b) this model is compared with experimental data. The model reproduces the correct trends here as well as for larger ensembles of islands such as in Figure 3, where it predicts that a population of unstable small islands will disappear to leave a relatively narrow distribution of larger islands.

Figure 4

In this simple model, islands have no preferred size, but instead grow or shrink at a rate determined by their environment. Islands continue to grow as long as smaller islands are available to supply adatoms, until the final configuration, one very large island, is reached. In contrast to the models described previously [12­14], there is no factor which tends to equalize island sizes; the narrow size distribution is the result of selection of the largest islands at the volume V_c. It is clear that for best uniformity, growth should be stopped just after the last pyramids have disappeared; further growth or annealing does not improve the island distribution.

Narrowing the island size distribution, even below the width shown in Figure 3(c), would be advantageous to the performance of real devices. One prediction from our model is that the distribution at later times can be improved only by controlling the initial nucleation of the islands. If all of the islands were to nucleate at the same time and be equally spaced, they should develop similarly and have a very narrow distribution. We have therefore looked into the factors which control island nucleation.

Any irregularity on the initial Si surface which changes the local sticking probability of Ge adatoms can influence island nucleation, and islands are known to nucleate preferentially on mesa edges [18], near clusters of dislocations [19], and near step edges [20]. However, we have found that the strain fields from individual buried dislocations provide the most convenient way of controlling island nucleation. This effect is shown in Figure 5, where a SiGe alloy was used to generate buried dislocations. Growth of the alloy above the critical thickness for dislocation formation [21] resulted in the creation of a network of misfit dislocations at the SiGe/Si interface (see inset in Figure 5). Further growth of pure Ge resulted in island growth in rows associated with the dislocations. Figure 5 shows that islands do not form directly above the dislocations but are offset to one side or the other.

Figure 5

The nucleation of these islands is influenced by the strain field of the dislocations, which extends up to the specimen surface. A calculation of the surface strain field due to a single dislocation³ showed that the region of maximum tensile strain, where we might expect Ge atoms to stick preferentially, is offset to one side of the dislocation; the islands formed in a row at the expected location (Figure 6). (The nucleation of islands at this location was not simply an effect of the additional surface step due to the dislocation. This is because the step density was very high, typically one step every 10 nm on the surface; therefore, one extra step was not significant.)

Figure 6

These experiments have therefore shown that it is possible to pattern single rows of islands along individual dislocations. A two-dimensional array of dislocations should allow islands to adopt a regular pattern, resulting in a structure which may perhaps be conducive to information storage. Techniques for patterning arrays of dislocations have already been developed [22], and the fabrication of devices requiring regular arrays of quantum dots should therefore be feasible.

In summary, real-time observations of island self-assembly in Ge on Si(001) have led to the development of a simple model which can explain the dynamic changes in island population observed during growth. The presence of two island shapes leads to surprisingly complex kinetics, and since the island evolution occurs over such short time scales, it would be difficult to understand the growth process without carrying out in situ observations. From these results, we can suggest ways to approach the important problems of size control and patterning of quantum dots. In terms of size control, growth of a uniform group of islands is possible owing to the selection effect at the volume V_c and is best achieved by stopping growth just after the smaller islands have disappeared. For patterning, the strain fields of individual dislocations can influence island nucleation strongly enough to enable us to place islands in particular areas of the surface. Taken together, these results show that islands formed by strained-layer epitaxy may play a significant role in the development of quantum dot devices.

Phase transitions in TiSi₂

Figure 7

The contact scheme described above is complex but works well for circuit designs with contact areas greater than about 1 µm². However, the miniaturization of components in integrated circuit design has reduced the area available for such contacts and led to an unexpected problem: In very small areas, the conversion of C49 TiSi₂ to C54 TiSi₂ is retarded, occurring only at temperatures above ~900°C [23]. The phase transition occurs unreliably during the 850­900°C anneal, leading to contacts with widely varying resistivities. This problem has slowed progress in the development of future generations of circuits with TiSi₂ contact layers, and it is therefore very important to understand why the phase transition is inhibited in small areas, and how it can be encouraged to occur at lower temperatures.

As might be expected for such an important problem, several analytical techniques have been applied to the study of the C49­C54 phase transition [24], and X-ray diffraction and ex situ microscopy studies have revealed some important features of this phenomenon. Grain sizes in the C49 and C54 phases differ by about an order of magnitude, with C49 grains around 100 nm in diameter transforming to C54 grains more than 1 µm across. The very large size of C54 grains suggests that the nucleation of the C54 phase is difficult, and the nucleation sites for C54 have not yet been unambiguously identified. Further studies have shown that alloying Ti with other metals, such as Mo or Nb, before permitting it to react with Si can reduce the C54 grain size and improve the uniformity of the reaction in small regions. Synchrotron X-ray experiments following the progress of the phase transition in blanket and patterned films have revealed subtle changes in reaction temperatures and film texture which reflect the underlying mechanism of the effect of these alloying elements [25].

However, one important factor which has not been addressed in these studies is the effect of impurities other than alloying metals in the films. Because Ti is such an efficient gettering agent, sputtered films of Ti contain substantial amounts of oxygen, and even relatively clean films oxidize further between growth and annealing. This effect is particularly important for the thinnest films of most interest in integrated circuit manufacturing. We have therefore used the in situ cleaning and evaporation capabilities of the UHV­TEM to examine the effect of oxygen impurities on the TiSi₂ phase transition. We have compared the transition in clean, UHV-deposited Ti with the transition in sputtered films which contained appreciable amounts of oxygen. The differences in reaction kinetics were striking and showed that the presence of oxygen at grain boundaries in the C49 phase causes a significant reduction in the rate of the C49­C54 phase transition.

To observe the transition in situ, a clean (001) Si surface was formed by repeated flashing of a thinned Si specimen to 1300°C, and a 10-nm-thick Ti film was deposited at room temperature at a pressure of less than 5 × 10^­9 Torr. The specimen was then heated progressively in the microscope to 850°C. The phases formed were identified from images and diffraction patterns, and the reaction kinetics were recorded at video rate. A similar heating sequence was applied to a 10-nm-thick Ti film on Si(001) which had been deposited by sputtering ex situ; the specimen was then thinned to electron transparency by etching from the back surface.

Figure 8 shows the phases formed during the heating of the UHV-deposited Ti film. As-deposited, the Ti film had a grain size of around 10 nm [Figure 8(a)], and annealing at low temperatures resulted in the formation of a small-grained TiSi phase (not shown). The C49 TiSi₂ phase appeared after heating to 700°C, while the C54 phase appeared at 850°C.

Figure 8

The phases formed during heating of the ex-situ-sputtered film appear indistinguishable from those for the UHV-deposited film, with similar grain sizes and diffraction patterns. However, the kinetics of the C49­C54 phase transformation were quite different. In Figure 9 it can be seen that the transition occurred smoothly in the UHV-deposited film, with a large C54 grain steadily consuming the smaller C49 grains. The interface velocity was independent of the orientation of the C49 grains. In contrast, the sputtered film underwent a very jerky transformation, pinning at grain boundaries in the C49 structure, then moving rapidly across C49 grains. We attribute the differences in reaction kinetics to oxygen segregation at the grain boundaries. Since C49 and C54 have the same stoichiometry, the reaction in the UHV-deposited film could occur relatively easily without any need for long-range diffusion. However, the reduction in the number of grain boundaries which also occurred during the transformation requires diffusion of any grain-boundary segregants to other sites on the surface or interface. It is this process which appeared to reduce the reaction rate in the ex-situ-sputtered film.

Figure 9

This result is significant in our attempts to enhance the C49­C54 reaction rate during integrated circuit processing. For a rapid transformation, low oxygen concentrations are preferable. However, it is also advantageous to increase the nucleation density of the C54 phase so that individual reaction fronts have less distance to cover before the reaction is completed. Since the C54 nucleation density can be increased by alloying, we are now conducting further experiments to determine how oxygen impurities affect the transformation kinetics in alloyed films.

Studying the liquid/solid interface by in situ microscopy

Several phenomena are unique to the electroplating process used in integrated circuit fabrication. These include the effect of the underlying “seed layer” on grain nucleation and orientation, and a “transient grain growth” process in which grain size increases, even at room temperature, in the minutes or hours after deposition. Observations made in real time during and immediately after deposition would be helpful in understanding these phenomena. We are therefore developing a liquid cell in which electroplating, or other liquid processes, can be reproduced while under TEM observation.

The cell, shown schematically in Figure 10, is designed to fit into a single-tilt holder which is of standard design except for the addition of electrical connections. It is made up of two silicon-nitride-covered Si wafers which have been etched to form small windows. The wafers are glued together face to face so that these “imaging windows” are aligned, while a spacer layer keeps the windows a fixed distance apart. Two larger windows are also etched into the upper wafer, and the nitride is removed from these windows. The two etched volumes act as reservoirs for the liquid, so that a drop of liquid can be placed in one reservoir and can flow by capillary action across the imaging window. After liquid is introduced, the reservoirs are sealed by gluing sapphire squares on top of the cell.

Figure 10

To study electroplating, electrical contacts are patterned on the wafers before the cell is assembled (Figure 10), and a copper sulphate/sulphuric acid solution is used as the electrolyte. It is intended that plating will take place onto an anode which is visible through the imaging window. Calculations suggest that application of a potential of around 2 V will allow a current density of 10^­6 A/mm² to flow, depositing 200 nm of Cu in several minutes. Reversing the polarity allows the copper to be dissolved off the anode so that the experiment can be repeated. It is important to note that the volumes of copper which are to be plated are sufficiently small that the finite amount of electrolyte present is not significant.

Although the full system is still undergoing testing, we have already shown that cells of this design can hold liquids in the vacuum environment of the microscope. However, due to curvature of the SiN, the total specimen thickness (liquid, metal, and SiN) is much greater than desirable for imaging. To improve the image quality, we have recently added an energy filter to the microscope.⁴ This improves the image contrast significantly by removing electrons which have undergone inelastic scattering, allowing only elastically scattered electrons to contribute to the image. The development of the liquid cell should enable us to study many phenomena other than electrodeposition, including switching in liquid crystals and crystal growth from a solution.

Conclusions

Acknowledgment

References

Footnotes

Received July 12, 1999; accepted for publication October 15, 1999