According to the mTLM extraction procedure, the Rf should be obtained first, and then the ρce could be extracted from the Rf-1/Wc plot. To extrapolate the Rf, the Rtotal-Lc plot is necessary and shown in Fig.3-6 and Fig.3-7. In Fig.3-6 with Wc = 1 μm and Lc = 2 μm, the extrapolated Rf value could range from -46 to 106 Ω if
61
different group of devices is chosen due to process variation. In addition to the uncertainty of the extrapolated Rf value, the unreasonable negative Rf is also likely to be observed. Intrinsically the extrapolation method could be one reason for the inaccurate Rf [34,40,41]; moreover, the condition that the Rs is considerably higher than the Rf would intensify the Rs variation dependence on the Rf and deviate the extrapolated results, i.e., the error of the Rs could play a major role during the extrapolation instead of the Rf if the Rs is relative larger enough than the Rf, which is consistent with the simulation results mentioned in the previous chapter and the literature [34]. Fig.3-7(a) gives the mean value of the Rtotal for each Ld, and the linear relationship between Rtotal and Ld could be seen consequently. Hence, it is noted that basically the mTLM method suffers from both the process variation and the intrinsic error by using the extrapolation method; however, by means of the law of large numbers in probability theory mentioned in previous chapter, average of sufficient experimental data could diminish the inaccuracy and correct the extrapolation, which is a feasible way to realize the mTLM method.
Next, in order to extract the ρce from the Rf-1/Wc plot, the Rf values extrapolated with Wc = 1.5 μm and 2 μm are obtained and shown in Fig.3-7(b) and Fig.3-7(c), respectively. The Rf-1/Wc plot is illustrated in Fig.3-8(a), whose slope gives the ρce, and the ρce = 2.9×10-7 Ω-cm2 for Lc = 2 μm consequently. Similarly, following the above procedure, the ρce = 9.8×10-7 Ω-cm2 for Lc = 1 μm is also extracted and shown in Fig.3-8(b).
For the mTLM method, issues mentioned in section 2-5-4 are also discussed in this section. First of all, the process dependence on the ρce extraction is obvious. It could be observed that the ρce of Lc = 1 μm is different from that of Lc = 2 μm by about two times; however, this noticeable difference between Lc = 1 μm or 2 μm for ρc in the range of 10-7 Ω-cm2 should not exist according to the simulation results
62
remarked in Fig.2-13 in section 2-5-3. This implied that the process variation truly exists and makes a little difference to each test structures even though test structures are fabricated in accordance with an identical process flow. The dispersive data shown in Fig.3-6 could confirm the influence of the process variation. Comparing with the simulation results in Fig.2-16(d), the process dependence could vary the ρce in a much wider range, 10-3 to 20 times for instance, and therefore leads the extraction to be worthless. Fortunately, after averaging Rtotal values for each Lc, the dependence of the process variation could be lessened, and therefore the extraction results of the mTLM method would be more trustworthy. As for the recessed contact, it was expected that the ρce would increase with the recession depth in a few times, illustrated in Fig.2-19(a) and explained in section 2-5-4. As the process dependence is significant, the influence of the recession may not be observed in this work; while the process dependence is diminished by averaging the data, the influence of the recession would emerge. It is supposed that the process dependence is reduced since the average Rtotal
and the Lc is in a linear relationship shown in Fig.3-7. Hence, the impact of the recessed contact would appear, resulting in an overestimate of the ρce. Moreover, since the extraction would be more accurate with larger Lc, it could be inferred that the true ρc would be lower than 2.9×10-7 Ω-cm2.
The Rs could be also extracted by the mTLM method, and the extracted Rs is 158.1 Ω/□ with Lc = 1 μm and 157.1 Ω/□ with Lc = 2 μm. Compared with the Rs = 140.6 Ω/□ measured by the four-terminal structure, since both methods are sensitive to the pattern distortion after the lithography, their results seems to be compatible.
To inspect the cross-sectional contact region of test structures, TEM was performed and the micrograph is shown in Fig.3-9. It could be observed that the NiSi layer is smaller than the actual contact size, and some unexpected voids appear between the NiSi layer. Also, there is an additional thin interlayer between NiSi and Pt
63
layers, which could be attributed to the wet etching process with the SPM solution for the unreacted metal removal [70]. In this process flow, a one-step annealing for silicidation was used, and hence two phases, Ni2Si and NiSi, would appear at the contact. After the selective etching with the SPM solution was performed, the follow-up DHF dipping may remove the interlayer partially and resulted in some gaps.
The last sintering step is supposed to make the Ni2Si transformed into the NiSi, and the excess Ni would tend to outdiffuse to the Pt layer through the gaps [71,72].
Consequently, the voids were left. Briefly, a smaller actual Ac is caused by the loss of the NiSi layer, and if the contact size could be correct, the ρce could be lower for both CBKR and mTLM methods.