Physical and statistical models have been used extensively to understand and improve electromigration and stress migration induced damage in narrow metal interconnects. As devices continue to shrink, and the interconnect delay becomes a more important fraction of overall circuit delay, the semiconductor industry has begun to make several important changes to the ‘standard’ Al-based interconnect technologies, including (i) a move from metal deposition and metal etch to metal deposition into damascene and dual-damascene structures, (ii) a move from Al-based to Cu-based metallizations, and (iii) a move from oxide to low-k dielectrics between the metal interconnects. The dominant electromigration and stress migration failure mechanisms associated with these advanced interconnect structures are typically different from the failure mechanisms associated with standard Al-based interconnects. In this work, physical and statistical models are used in conjunction with reliability data to compare the failure mechanisms in standard, Al-based interconnects with Cu-based, damascene interconnects. Methods are discussed for determining the dominant failure mechanisms and improving reliability in both types of interconnect.

A status report is given on a comprehensive modeling project aimed at predicting failure time distributions of interconnect lines. We discuss our novel approach to calculate the evolution of stresses in lines using elastic response functions. It is argued that this approach makes it possible to model the stress and damage evolution in a large ensemble of lines efficiently so that statistically meaningful failure time distributions can be generated.

The elastic response functions enable us also to derive a generalized Korhonen equation which includes the effects of mass transport at remote locations. Basic features of this equation are demonstrated with a one-dimensional implementation and its results are compared with the classical Korhonen [8, 9, 10] model.

We have simulated the development of grain structures in polycrystalline films with lognormally distributed grain sizes, and carried out extensive characterization of grain cluster and bamboo cluster statistics. These statistics were characterized as a function of line-width to initial-grain-size ratios in as-patterned strips, and in strips in which we have simulated post-patterning grain structure evolution resulting from annealing. Among the important findings is that polygranular and bamboo cluster-length distributions for as-patterned lines are best fit by Weibull distribution functions instead of lognormal or exponential functions, as is often assumed. We report analytic formulae describing grain structure statistics that can be used in reliability calculations and simulations. We have carried out structure-sensitive electromigration simulations to show that the predicted failure statistics are essentially the same for grain structures generated using grain growth simulations and grain structures generated using analytic models. Analytic models provide computationally efficient means of determining the lifetime variations associated with grain-structure variations.

A theoretical analysis is presented of the morphological evolution of transgranular voids in metallic thin films driven by capillarity, stress, and electromigration. This highly nonlinear and complex dynamics is analyzed through self-consistent numerical simulations based on a two-dimensional phenomenological model of a passivated interconnect. The analysis follows a systematic exploration of a six-dimensional parameter space determined by the strengths of the electric and stress fields, surface diffusion anisotropy, and the size of the void. Our simulation results reveal a very rich nonlinear dynamical picture of void morphological evolution mechanism, including faceting, wedge void and faceted slit formation through a facet selection process, and soliton-like surface wave propagation. These results are in very good agreement with experimental observations and have important implications for interconnect failure.

Future semiconductor technologies with structural dimensions less than 0.35 µm and four or more metallization layers impede the realization of the reliability requirements due to their complexity and increasing operating conditions (>5 mA/µm2). Therefore it is necessary to provide reliability appraisal methods which ensure the operation lifetimes (up to 30 years) required by the market.

The effects of metal Joule heating in interconnects were studied numerically. Particular attention is devoted to the multilevel nature of interconnects in modem microelectronic devices. Heat conduction analyses were carried out to quantify the temperature rise in structures composed of various levels of metal lines under different electric current densities. Two types of metallization (aluminum and copper) and two types of interlevel dielectric (silicon dioxide and polyimide) were considered. It was found that increasing the total number of metal level and/or switching the dielectric from silicon dioxide to polymer-based low-k dielectrics can cause substantial temperature increases, pointing out that interconnect Joule heating can become a major reliability threat in future applications. Thermal stresses induced by the nonuniform temperature field were also analyzed.

Microstructure plays a key role in reliability of interconnects. However, a quantitative correlation between microstructure and reliability of interconnects is lacking. This paper develops a statistical description of effective diffusivity of interconnects by considering grain boundaries, bulk lattice and line interface diffusion, and introduces a microstructure-based electromigration lifetime prediction model for normal failures of interconnects. Further, the probability of early failure of interconnects due to a severe microstructural inhomogeneity is calculated. Using the sub-population failure model, the mean time to failure of interconnects, with a range of microstructures from polycrystalline to hyper-bamboo, are simulated. The reported increase in lifetimes of wide polycrystalline lines with increasing w/d is mainly ascribed to reservoir effects of bonding pads in test structures. Reported lifetime prediction model may be used as guidance for circuit design and accelerated testing.

We describe a newly-developed defect generation mechanism, namely the grain boundary Frenkel pair (GBFP) model, and corresponding diffusion mechanisms during electromigration developed using atomic molecular statics (MS), Monte Carlo (MC), and molecular dynamics (MD) simulation techniques in Al and Al-Cu. We contend that large numbers of interstitials and vacancies exist at grain boundaries and both contribute to mass transport. Cu preferentially segregates to the interstitial sites at grain boundaries via a Frenkel pair generation process and reduces the overall grain boundary diffusivity due to the strong binding in the Al-Cu dimer. Predictions from our models are in excellent agreement with available experimental data and observations.

The direct influence of the applied electric field on ions of lattice, as a whole, was considered. This influence induces the strains and stresses and its gradients. In turn, this additional stresses may induce the vacancy (ion) diffusion flux. It is shown that this flux coincides in the direction and is comparable in magnitude with the electromigration flux induced by the electron wind.

In this paper, we study the influence of the line width on the mechanical behavior of narrow unpassivated lines of pure Al and Al with 0.5 wt.-% Cu. The stress/temperature behavior during thermal cycling is investigated using x-ray diffractometry. Our results suggest that it is possible to distinguish between two temperature regimes in which different deformation mechanisms take place. At low temperatures, the narrowest lines sustain larger stresses than the continuous films, whereas at high temperatures the opposite behavior was found. Further, iso-thermal stress relaxation measurements on pure Al lines after various heat treatments were performed. These results indicate the existence of a threshold stress below which no deformation of the lines occurs.

The thin film reaction between Ti and Al-0.5%Cu to form TiAl3 is common in the microelectronics industry. In this paper the stress changes in Al-0.5%Cu films at elevated temperatures during the reaction are measured. The changes are measured in blanket films as well as in passivated interconnect lines. Results show that in blanket films the Al-0.5%Cu does not experience any stress change due to the reaction. However in passivated lines, where the layers are not allowed to relax in the normal direction, tensile stresses build up in the Al-0.5%Cu due to the volume shrinkage that happens when these films react.

Sensitive measurements of electromigration-induced resistance changes in metallization interconnects have been previously interpreted as due to precipitation effects, vacancy concentration changes and local stress changes (due to the piezoresistance effect). The latter have the potential for characterizing the early stages of electromigration-induced diffusive phenomena prior to larger scale effects such as voiding and subsequent localized joule heating which seriously complicate the interpretation of small resistance changes. However the utility of such “early" measurements is also hindered by lack of an accurate knowledge of the stress coefficient of resistivity as a proper characterization of the change in strain state of the Al. We present measurements of the effects of volumetric strain on the resistivity of Al interconnects where the strain is induced by thermal expansion mismatch between the Al and surrounding passivations. The piezoresistance effect is characterized by properly accounting for the degree of interconnect constraint (and volumetric strain) induced by thermal expansion mismatch as a function of temperature for both passivated and unpassivated lines. Sensitive interconnect resistance versus temperature measurements for differently constrained interconnects (with different volumetric strains) thus allows for the measurement of the piezoresistance effect. Other effects such as solute or vacancy concentration changes with temperature are minimized since the measurements are performed by cooling the passivated and unpassivated lines from room temperature to approximately 70K, rather than by significant heating above room temperature. We determine the piezoresistance coefficient, defined as dρ/dɛv, where ρ = resistivity and ɛv = volumetric strain, to be approximately 1.5−5 Ω-cm, in rough agreement with previous work. The interpretation of sensitive resistance measurements as volumetric strains in isolated sections of interconnects undergoing electromigration-induced diffusion processes is described for specialized test structures.

A new test structure for electromigration failure analysis of via-interconnect metallization schemes was developed. Multi-via/interconnect chain arrays were used in a parallel/serial testing arrangement. Due to improved statistical sampling, the multi-via test structures enabled measurement of drift velocity phenomena at via/interconnect interfaces with a very limited number of test devices per stressing condition, and without reducing the signal to noise ratio. Furthermore, realistic, production-type metallization schemes can be characterized without usually required “traditional” drift velocity-type metal stacks incorporating highly resistive shunt-layers. Critical length effects were investigated as a function of line length and current density. At interconnect lengths close to the critical Blech-length, resistance saturation effects were encountered and used for calculations of the critical current density-length product, (jl)*. Furthermore, a new resistance saturation model was developed and found to correlate well with the acquired data. Critical length effects can accurately be measured and used for length-dependent interconnect performance evaluation on the device-level.

This paper reviews experimental data on stress- and electromigration (EM)-induced voiding in passivated Al- and AlSiCu-lines under different experimental conditions (thermal cycling, iso-thermal annealing without and with EM-current). The data are deduced from measured changes of average metal strain or stress and from in-situ observations in a transmission electron microscope. There are excellent correlations between corresponding data sets showing stress induced voiding during cycling and isothermal annealing without EM-current, but nearly no changes in stresses or void volume during EM-testing. The implications of these results on EM-modelling are discussed.

The electromigration behavior of pure Al lines passivated with oxides of different thicknesses and passivation deposition temperatures was studied. The initial hydrostatic stress states of the passivated Al lines were modeled with finite element modeling (FEM), and, when possible, measured with X-ray diffraction. Conventional wafer-level electromigration tests showed a clear passivation thickness effect, but no detectable effect of initial stress on electromigration lifetimes. Increasing the passivation thickness increased the electromigration lifetimes, which has been observed by other researchers. However, in a sample set where the Al lines were covered with a thin (0. 1µm) oxide layer, the lifetimes were much longer than expected. Differences in the damage morphology and the failure mechanism between the thin and thicker oxides accounted for this unexpected result.

This work shows that shear failure in the metallic power line on the IC chip, which mainly takes place after plastic packaging, is always followed by mechanical damage in its overlayer. Thus, the topological feature of the metallic overlayer is important in defining shear damage in the metal line. On the other hand, since open failure in the metal line associated with stress migration is caused by tensile stress resulting from thermal mismatch with its overlayer, the physical properties of the overlayer plays an important role in determining open failure in the metal line. The FEM simulation shows how the tensile stress distribution along a metal line varies with the physical properties of its overlayer. The metallic overlayer with an initial coating of a ductile oxide layer was estimated to be a very effective way of minimizing tensile stress induced to the metal line. On the other side, shear damage of the metal line was found to be highly sensitive to the chip location or the package structure. The numerical calculation also indicates that shear stress applied to the passivated metal line, which reveals a maximum value at the chip comer, can be considerably reduced by chip premolding overcoating with ductile photosensitive polyimide.

Photoresist (such as PMMA) thin films are among the most important materials in semiconductor fabrication processes. Photolithography exclusively uses photoresist to construct the circuit pattern. The quality of a pattern relies on the uniformity of a photoresist layer. However, thermal expansion mismatch between the photoresist and silicon substrate during the subsequent thermal cycling process inevitably produces undesirable residual strains/stresses in the resist layer. Such residual fields can deteriorate the desired quality of the final pattern. A recently developed method is used here to record full-field strain information in the thin resist film.

Traditional methods such as wafer curvature measurement or x-ray diffraction (for crystalline films) have not been effective for measuring the full-field strains. However, using a deposited dot array in the interface between the film and substrate, one can measure the strains in the thin film, full-field, by analyzing the displacements of individual dots. A biquadratic surface model is employed to fit these dot pattern displacements. An IDL (Interactive Data Language, Research Systems Inc.) program is utilized repeatedly to accurately locate the positions of the individual dots and transform that information into displacement and strain data by differentiating the biquadratic surface. This study successfully demonstrates a straightforward approach to determine the residual strains, full-field, in photoresist thin films.

It is well known that the addition of Cu improves the electromigration resistance of Al conductor lines. However, the underlying mechanisms are not entirely understood. As Mg has also been found to have a beneficial effect, a thorough comparison between these two alloying elements may result in new insights. In this work, we present results of drift velocity experiments on Al-2wt%Cu and A1-3wt%Mg, where the edge drift was measured in-situ as a function of temperature. Additionally, the concentration profile of either alloying element was measured as a function of testing time using an electron probe micro-analysis technique. For both alloys an incubation time for stripe drift was observed. The activation energies of edge drift after this depletion were determined to be 0.83 eV and 1.01 eV for Al-Cu and Al-Mg, respectively. Both activation energies seem to agree with the activation energy of electromigration of the alloying element rather than the activation energy of drift in pure Al, which was determined separately to be 0.5 eV.

To investigate why the formation of electromigration damage in Al lines alloyed with both 0.1 at.% Ni and 0.1 at.% Cr is so effectively suppressed we studied electromigration induced changes in the electrical resistance of short lines. The length of the lines is 4, 6, 8, or 12 µm, the width is 4 µm and the film thickness is 0.3 µm. As previously reported for pure Al lines, depending on current density and sample length, the induced resistance changes fully recover or do not recover after DC stressing for a few hours. The recoverable resistance changes correspond to the build-up or relaxation of mechanical stress and the permanent resistance changes to the formation of a void or hillock. Electron microprobe analysis and resistance measurements after prolonged current stressing provided no evidence for changes in local solute concentration. Therefore diffusion of Al along grain boundaries is regarded as the dominating diffusion process. The activation energy and the pre-exponential factor of this process are determined from the temperature dependence of the relaxation time of the recoverable resistance changes. The activation energy of this process is 0.84±0.03 eV and the pre-exponential factor δD0 = (1.8±0.8)×10−8 cm3/s. At 175 °C the resulting diffusion coefficient is about a factor 70 smaller than that of pure Al. The transition from recoverable to permanent resistance changes is characterized by a constant critical current density -sample length product. It is found that the value of this product for the Al-Ni-Cr lines is two times larger than the value found for pure Al lines. This means that the threshold current density at given sample length for the formation of hillocks or voids is two times larger. Both the slower solvent diffusion and the higher threshold current density contribute positively to a long electromigration lifetime.

The electromigration characteristics of CVD-Cu films were compared to those of PVDCu films. The CVD-Cu films were deposited using metal-organic precursor of (hfac)Cu(VTMS) and the PVD-Cu films were deposited by evaporation of a 99.995 % Cu metal on TiN and SiO2 substrates. To test EM-resistance, Cu film was patterned into the test structure of multi-lines with 20∼30 µm width and 2500 µm length per one line by a photolithography technique. These test lines had polycrystalline structures composed of much smaller grains than line width and thickness. An accelerated EM test was carried out in the vacuum of 10−6 Torr because the test in N2 ambient had lead to a resistance increase mainly due to oxidation of Cu.

The CVD-Cu lines had shorter lifetime than PVD-Cu and the activation energy of electromigration was 0.754 eV for the CVD-Cu on SiO2, which means the grain boundary diffusion as the EM failure mechanism. The CVD-Cu films had larger grains and poorer continuity between grains than the PVD-Cu films so the resistivity of CVD-Cu films (3.20±0.21 µΩcm) was higher than that of PVD-Cu films (2.00±0.06 µΩcm) having a value of near bulk resistivity. The higher resistivity induces the higher driving force of EM. Moreover the PVD-Cu films had a huge domain structure, which might act as a larger grain than CVD-Cu grain to result in the longer life-time and the surface diffusion mechanism with the activation energy of 0.244 eV.