Dominant electromigration failures in copper/low-k interconnects
When multiple electromigration failure modes exhibit bimodality it may be time to question void formation. Shou-Chung Lee and Anthony S. Oates of TSMC report on the findings of electromigration-induced void morphologies.
Via connections between metal layers in copper/low-k dual-damascene interconnect structures are a common source of reliability problems. Shou-Chung Lee and Anthony S. Oates of TSMC report on investigations of electromigration-induced void morphologies that dominate the reliability of Cu/low-k dual damascene vias using multi-link structures to enhance early failure sensitivity.
Dual damascene copper vias exhibit multiple electromigration (EM) failure modes [1, 2]. Consequently, bimodality is often observed in Cu electromigration failure distributions, particularly when electromigration-induced voids form in the upper levels of vias. These voids can occur inside vias and in the attached metal stripe. Broad failure distributions are often observed, however, for void formation under vias. Several failure modes have been observed, with the failure time appearing to depend on the void shape and location. A clear understanding of the fundamental, dominant failure modes for low-k dielectric structures has yet to emerge due to the influence of early failures related to processing issues and defects. Here, based on work first presented elsewhere [3], we show that voids formed directly under vias present the most serious issue, and pose a fundamental limitation to Cu via reliability. Moreover, we observe that early failure distributions can be present at low percentile levels, necessitating the use of testing techniques with sensitivity below 1%.
Experimental details
Electromigration-induced void studies were carried out between the 0.13µm to 65nm
nodes. In all cases SiOC based inter-level dielectrics were used with damascene
structures, which were defined by dry etching of the dielectric layer, followed by deposition of a Ta-based trench liner. Cu interconnects were defined by standard electro-plating and CMP planarisation, and were passivated with a dielectric barrier.
Via failure modes were evaluated using a variety of test structures to study either “downstream” electron flow from the upper to the lower level (V1M1) or “up-stream” flow in the opposite direction (V1M2). The width of the metal stripe attached to the vias was varied typically between the minimum for the technology node, wmin, and the maximum geometric width that can sustain a single via across the stripe width, usually ~3 wmin. The structure schematics are shown in Figure 1. In practice, stripe widths varied between 0.1µm and 0.4µm, while via diameters ranged from 0.1µm to 0.13µm. In all cases the line length between vias in the link structure was 250µm.
The structures were designed with the via diameter less than the surrounding stripe width, so electrical redundancy between the vias and the trench liner was not anticipated. The failure criterion in all cases corresponded to the initial resistance jump for both V1M1 and V1M2 structures.
While single link structures are useful to study basic voiding mechanisms, as will become evident, increased early failure sensitivity is needed to evaluate Cu via reliability. To achieve high resolution to early failure populations we use multi-link structures consisting of serial chains of identical via units. Multilink chains significantly increase the sample size for EM experiments, allowing early failure distributions to be observed at much lower percentiles than for single via structures. For multi-link structures the number of identical V1M1 or V1M2 links in series varied between 10 and 2400. For link numbers less than 100, conventional Kelvin resistance measurements are sufficient to detect all electromigration-induced resistance increases. For larger link numbers, however, a high sensitivity Wheatstone bridge was used to ensure that changes in the large chains are detected accurately. Each of the four sections of the Wheatstone bridge consisted of a total of 600 wmin links. The bridge imbalance voltage was monitored with a resolution of 1µV, implying sensitivity to changes of about 1? for the test
structure.
Multi-link electromigration failure
The analysis of the failure distributions of multi-link structures is based on the statistical weak-link concept. The multi-link structure fails when one link does. Consequently, the failure distribution of a multi-link structure, FN(t), is related to that for a single link, F1(t), by:
F N(t) = 1 – (1 – F1(t)) N
where N is the number of links. The failure distribution of a single link can, therefore, be obtained from inversion:
F1(t) = 1 – (1 – FN(t))1 /N
A Monte-Carlo simulation on a 50-link distribution showed results consistent with a test single-link distribution. This plot also showed that much lower percentile levels can be measured using the multi-link technique.
If, however, the multi-link distribution is bimodal a different deconvolution procedure must be used to separate early and late failures for single vias. When early failure populations are present, the single link and multi-link distributions can be expressed as:
F1(t) = xF1E(t) + (1 – x) F1L(t)
FN(t) = p FNE(t) + (1 – p) FNL(t)
where F1E(t), F1L(t) and FNE(t), FNL(t) are the cumulative failure distributions of the early and late single link and multilink distributions at time t, respectively. The variables x and p represent the early single- and multi-link failure probabilities, respectively. These parameters are not independent:
p= 1 – (1– x) N
Assuming F1L(t) ˜ 0 and FNL(t) ˜ 0 at early times, the singlevia early distribution can be obtained from the measured bimodal multi-link early failure distribution [3] by:
F1E( t) = [1 – (1 – pFNE( t))1 /N over 1 – (1 – p)1 /N
Similarly assuming F1E( t2) ˜ 1 and FNE( t2) ˜ 1 at late times, gives:
F1L( t) = 1 (1 - FNL( t))1 /N
The assumption is that the early and late distributions do not overlap so that the early failure probabilities xand pcan be clearly determined from the bimodal distributions. This is usually the case for Cu interconnects where early failures usually occur at much earlier times than late failures [4]. We confirmed the applicability of the two deconvolution equations to bimodal multi-link distributions using Monte Carlo simulations.
Upper to lower electron flow
Stressing of via structures with electron flow from the upper to lower metal (“down-stream”) leads to either abrupt large or progressive monotonic resistance increases with stress time, irrespective of the lower stripe width. Examination of stressed vias reveals that progressive resistance increases are associated with voids that are either:
- under the via, but displaced along the line-length (Figure 2(a)), or
- in the line attached to the via (Figure 2(b)).
In both cases, the trench liner maintains electrical redundancy. The abrupt resistance changes, however, originate only from voids directly under vias, where the void is shaped as a narrow slit (Figure 2(c)). The only variation in this failure mode appears to be the void extension in front of and a smaller variation in the depth below the via. The images in Figure 2 are collected over several process generations and are representative of electromigration failures in V1M1 structures. Both the abrupt and progressive increase modes are observed, irrespective of fabrication process details associated with the Cu trench or via. These two modes thus appear to represent the fundamental failure modes for Cu/low-k vias.
For optimized process technologies, small sample (~20–30), single-via experiments frequently exhibit both types of resistance increase at similar stress times (Figure 3(a)). While there is variability in the relative proportion of the two types of resistance increase even after process optimization, both modes appear to be always present. With single via experiments, the difference between the failure distributions of both modes is difficult to unambiguously interpret, but these measurements tend to indicate that the lognormal dispersion parameter, s, associated with the progressive increase mode is ~0.4 compared with ~0.8 for the abrupt increase mode (Figure 4). The larger s of the abrupt increase mode implies that it dominates at low failure percentiles, and thus determines via reliability. We unambiguously confirm these observations using multi-link structures that show a clear distinction between the modes (Figure 3(b)). It is clear that abrupt large resistance increases occur first. Despite the observation of two different fundamental types of void formation, slit-voids apparently determine via reliability in the down-stream direction. It is also possible to observe small progressive resistance increases of up to about 30 ? before the large abrupt jump. We believe this to be related to multiple, small randomly located voids in the 50-link structure.
Figure 5 shows an example of the failure time distribution associated with slit-voids, measured at wmin stripe width. Both single and multi-links with up to 2400 identical links have been used to determine failure at percentiles below 10-3%. We believe this is the first time that Cu via electromigration data has been reported at such low percentiles. These measurements have been performed with a current density that is sufficiently high to ensure that the critical current density has negligible influence on the failure distribution. Previous work showed that as stress current densities approach critical, electromigration failure can be distorted from the lognormal form, making interpretation difficult [5]. In the 2400-link structure, the measurements corresponded to resistance changes of 600–1000?, confirming that abrupt increases occur across the whole distribution.
The measured distribution shows deviation from the lognormal form at the lowest percentiles. Above about 0.1% the distribution has an effective s ~ 0.9, while below 0.1% it decreases to ~0.6. Nevertheless, these deviations are of limited practical impact since typical reliability requirements are in the range 0.1–0.01%, where the lognormal distribution is a good approximation. Because of the sensitivity of the multi-link technique, these percentiles can be directly measured, removing the need for extrapolations based on mathematical fits. Because the failure distributions associated with slit-voids tend to be broad, some care must be taken to ensure that the sample size is large enough to eliminate statistical artefacts. Figure 6(a) shows an example of an experimental failure distribution obtained with a relatively small experimental sample size of 11 structures, each with N= 50. An apparent bimodal distribution is observed, even though failure analysis clearly showed all voids to be slit type. Further measurements, increasing the sample, eliminated the apparent bimodality, leaving a characteristic broad single, lognormal distribution (Figure 6(b)).
The failure distributions of V1M1 multi-link structures over the range wmin to 3wmin exhibit similar s values, as expected, since the slit failure characteristics do not depend on the lower stripe width. Wider stripes, however, tend to show lower failure times at the same stripe current density. As-measured multi-link failure distributions for wmin and 3 wmin show ~2 x shorter failure times for the wider stripe compared to the narrow one.
We have observed this lower failure time for 3 wmin structures to be a general feature of slit-voids. We speculate that this difference in failure time is associated with the smaller void volume compared to the narrow stripe, perhaps from voids that do not extend across the stripe width.
Lower to upper electron flow
“Up-stream” electron-flow (i.e. from the lower to the upper metal level), produces only progressive, monotonic increases in resistance irrespective of the width of the upper stripe. While the resistance change appears to be similar for all samples, two types of voiding occur for up-stream electron flow. Voids form either inside vias (Figure 7(a)) or in the stripe attached to vias (Figure 7(b)). As discussed by others, voiding within vias is strongly process dependent, and can be suppressed by optimization of the via barrier and cleaning/etch processes [6]. This failure mode, therefore, does not present a fundamental limit to via reliability. It is possible, however, for failure times associated with this mode to be sufficiently low to limit via reliability below that of the slit mode. Voiding in the middle of the metal line attached to the vias produces significantly longer failure times compared to those within the via and thus does not directly impact practical reliability.
Because of the early failure potential associated with void formation inside vias, it is particularly important to characterize failure at low percentiles. The up-stream electron distribution is typically determined with a multi-link structure to increase early failure sensitivity. The single-via early failure level over the wmin to 3 wmin range is around 2%, a low level to characterize accurately without multi-links. Deconvolution of the bimodal multi-link failure distributions into single via distributions for wmin and 3 wmin (Figure 8) shows that the early failures associated with voids inside vias can exhibit an order of magnitude reduction in time compared to voiding in the stripes attached to vias. The late failure portion of each distribution is similar, which is consistent with the fact that the voiding responsible for these failures is identical for both widths, and is in the stripe attached to the vias. Early failures for 3 wmin, however, appear to be about a factor of 2 lower in failure time than wmin. These results again emphasize that reliability limitations may exist at wider stripe widths rather than at the minimum feature size. The origin of the failure time difference between 3 wmin and wmin for voiding inside vias is unclear at this time. Finite element simulations of the resistance changes associated with voiding inside vias indicate very small sensitivity to stripe width and to void position within the via. Comparing these early failures with the slit mode, we observe that voiding within vias, when present, may exhibit even lower failure times at low percentiles, and so may become the dominant mode in determining electromigration reliability. Elimination of this mode of voiding is, therefore, an important issue in ensuring high current carrying capability for Cu interconnects.
Applications of 2-level multi-link structures
Separate examination of voiding in the two possible electron flow directions gives a complete understanding of potential failures, but for reliability evaluation it is inefficient. For reliability prediction only the dominant failure mode is important. With the goal of identifying this dominant mode most effectively, we also investigated a multi-link structure combining V1M1 and V1M2 structures, where all lower and upper level links are identical (a 2-level multi-link, Figure 1(e)). This structure is capable of exhibiting all possible failures, but theoretically, it should always exhibit the dominant mode in accelerated testing. The failure distribution of a test structure that may fail in the upper or lower level, G( t), is given by:
G( t) = FN,V1M1( t) + (1 – FN,V1M1( t))FN,V1M2(t)
where FN,V1M1( t) and FN,V1M2( t) are the cumulative distribution functions of the V1M1 and V1M2 multi-link structures at time t, respectively. That this combined distribution always exhibits the dominant electromigration failure mode is easily seen: for V1M1 dominant (i.e. FN,V1M1( t) >> FN,V1M2( t)), G(t) ~ FN,V1M1( t), while, for V1M2 dominant, G(t) ~ FN,V1M2( t). The disadvantage of the 2-level (2L) structures is that physical identification of the dominant mode may not be evident from resistance change data and must be confirmed by physical analysis. Nevertheless, considerable time and resource savings are available when the primary focus is rapid assessment of overall reliability.
Experimentally, we confirm that 2L structures automatically identify the dominant electromigration failure mode in Figure 9, which compares measured single via V1M1 and V1M2 distributions with the deconvoluted distribution from 50-link 2L structures measured on the same wafer. There is excellent agreement between the 2L distribution and the 1-via V1M1 distribution, as expected, since the single-via V1M1 failure time is the lowest. Physical analysis demonstrates that the dominant failure mode for the 2L structure is again slit-like voids directly under the vias.
Authors:
Shou-Chung Lee, Anthony S. Oates, TSMC, Science-Based Industrial Park, Hsinchu, Taiwan 30077, R.O.C. scleec@tsmc.com
References
[1] Gill et al, Proceedings IEEE International Reliability Physics Symposium, 2002, pp.298-304.
[2] Li et al, Device and Materials Reliability, 2004, vol.4, No.1, pp.80-85.
[3] Lee and Oates, Proceedings IEEE International Reliability Physics Symposium, 2006, pp.107-114.
[4] Ogawa et al, Proceedings IEEE International Reliability Physics Symposium, 2001, pp.341-349.
[5] Huang and Oates, Proceedings IEEE International Interconnect Technology Conference, 2000, pp.208-210.
[6] Fischer et al, Proceedings IEEE International Interconnect Technology Conference, 2002, pp.139-141.