Case Study 7: Alloying Kinetics in Contact Metallisation for CMOS

Summary

Nickel silicides are widely used in making electrical contact to complementary metal-oxide-

semiconductor (CMOS) devices in advanced integrated circuits. They have been the preferred

contacting material since the 2006 “65 nm technology node”, partly because a self-aligned-silicide

(“SALICIDE” [1]) process is available. However, their high temperature behaviour is complex and,

crucially, improved by the presence of platinum.

Seminal information [2] on the influence of the Pt in limiting the formation of the undesirable NiSi2

phase during annealing was obtained kinetically using real-time RBS [3] (see Figure), where the very

large quantity of data resulting is analysed in real-time by an artificial neural network (Barradas &

Vieira, Phys. Rev. E, 2000 [4]).

This is an important example, but only one of many quantitative operando RBS [5] observations of

diffusion and phase separation during annealing of multilayer samples. Operando measurements

can be extraordinarily efficient compared to a conventional “cook and look” approach, and indeed

readily give details of processes that are hard (or impossible) to obtain conventionally.

Body

Clearly, Rutherford backscattering spectrometry (RBS) is ideal for observing silicidation processes,

since the metal signal is background-free and the depth resolution is well matched to the

application. Figure 2 shows conventional RBS of the important Ni:Pt system (Corni et al,

Appl.Surf.Sci., 1993 [6]).

The Ni:Pt/Si system is very complex however, and the presence of Pt has intricate effects on the

interplay between the two phases, Ni2Si and NiSi. This is most clear from the operando observation

of the complete silicidation process (Figure 1), a single spectrum from which is shown in Figure 3

where the segregation of Pt out of the di-nickel silicide and the onset of the formation of the

monosilicide can be seen.

Further work by Demeulemeester et al (J.Appl.Phys. 2013) [7] on this important system has shown

the sensitivity to the Pt content (see Figure 4). They successfully applied real-time RBS in

combination with artificial neural network analysis to disentangle the growth kinetics during the

complex growth of Ni(Pt) silicides and showed that activation energies can be extracted from a

single ramped real-time RBS measurement (see Figure 5).

Complementary work on the same system included real-time X-ray diffraction (XRD: Demeule-

meester et al, J.Appl.Phys. 2013 [8]) and atom probe tomography (APT: Mangelinck et al, Scripta

Materialia, 2010 [2]). RBS allows the phase formation and the redistribution to be followed in situ

and thus gives crucial information of the different steps of the redistribution. But RBS has relatively

poor depth resolution and, for these nanocrystalline materials, effective no lateral resolution. APT is

an ex situ but very high resolution tomographic method giving highly detailed structural information

(see Figure 6). Operando RBS gives a broad overview of the entire process, where APT follows

inhomogeneities at the nano-crystalline level. This is similar to the complementarity between RBS

and transmission electron microscopy (TEM).

Real-time RBS is a very powerful method of obtaining continuous information of phase formation

during annealing, which perhaps Theron et al (NIMB, 1998) [9] were the first to report for this

silicidation system. In-situ RBS has been used for a long time: see Rennie et al (Appl.Phys.Lett., 1986)

[10] for the photodissolution of silver in chalcogenide glasses, or Averback et al (Phys.Rev.B, 1983)

[11] on the radiation-induced segregation of nickel-silicon alloys for examples.

What is different here is the use of artificial neural networks (ANNs) to handle the huge datasets

produced quantitatively. The critical thing for an ANN is how to “train” it. A “training set” must be

constructed to allow the network to assign the proper weights to all its nodes so that the inputs elicit

the correct outputs. There are three essential things to understand about ANNs: (i) the ANN can

only interpolate within its training set, it cannot extrapolate out of it; (ii) the ANN knows no physics!

and (iii) the response of the ANN is effectively instantaneous. This is all described in detail by

Barradas & Vieira, (Phys. Rev. E, 2000) [4].

ANNs have had some bad press because it is easy to train them badly: they will always give an

answer, but the answer can be complete nonsense! However, in the cases we have described the

answer has been demonstrated to be reliable. Figure 7 shows the systematic way this is done for

real-time RBS.

The ANN used by Demeulemeester et al, (NIMB 2010) [5] is trained with a complex training set of

18000 spectra and a test set of 2000 spectra (generated automatically by the DataFurnace [12]

code). But all 63 spectra shown in Figure 7 are analysed for the thickness of the various phases (with

appropriate “roughness”) by the trained ANN in a fraction of a second. The ANN output is then fed

into DataFurnace as an initial “guess” at the structure implied by the spectrum, and fitted using a

proper physical model [13] (which is known very well!). Clearly, Figure 7 shows that the ANN was

very well trained! Equally clearly, the inversion of the RBS spectra (mathematically an ill-posed

problem [14]) can now be done entirely automatically! It is a huge advance to be able to fit very

large numbers of RBS spectra – and fit them effectively perfectly, thus extracting approximately all

their information – without any intervention by the analyst. This procedure works even for very

complicated RBS spectra, that have traditionally been a nightmare for the analyst. The analytical

difficulty is now transposed to constructing the proper training and test sets for the ANN.

There is even a suggestion that RBS without humans [15] is possible, proposed by Barradas et al

(Phys.Rev.E, 2002) [16]. They demonstrate that it is possible to create an algorithm based on ANNs,

“which is able, for a given system [in their case Ge-implanted Si], to optimize the experimental

conditions for each sample and then analyze the final spectrum collected. The algorithm is easily

extensible to other systems. Once this algorithm is implemented in a code connected to an

experimental setup with automated sample loading, this will lead to the performance of RBS

experiments entirely without the assistance of humans.” Importantly, Barradas et al here

demonstrate that ANNs can successfully be developed that are able to warn if a spectrum falls

outside the training set: this would be an essential safeguard for any automatic system.

However one views “RBS without humans”, what is certainly true is that IBA methods currently

depend very heavily on the expertise of the analysts. IBA is expensive! But there are surely

“routine” cases which ought to be amenable to automation, so that the price per analysed spectrum

is reduced? This works says that indeed there are, and moreover that we know how to treat them!

On a separate issue, this work is limited as it stands to annealing only to about 600°C, because hotter

samples will necessarily generate higher energy photons, and the semiconductor detectors are light-

sensitive. But we should note that there is a completely different detector technology impervious to

sample temperature. Muller et al (NIMB, 2011) [17] has described a very simple gas ionisation

detector with a performance comparable to the semiconductor detectors for He-RBS. These devices

have greatly enhanced energy resolution for ions heavier than He, operate equally well in the light

or at high temperatures, do not suffer from beam damage, and can support very high count rates.

They could transform this application, and IBA in general.

Keywords

Real-time, Artificial Neural Network, ANN, silicide, SALICIDE, CMOS, annealing, operando, automatic

analysis, gas ionisation detector

Thin Film Modification Methods

Sputtering, ion implantation

Complementary Analytical Methods

RBS, XRD, APT (atom probe tomography)

Cited Literature

[1] P.S. Lee, K.L. Pey, D. Mangelinck, J. Ding, D.Z. Chi, L. Chan, New salicidation technology with Ni(Pt) alloy for MOSFETs, IEEE Electron Device Letters, 22 (2001) 568-570; DOI: 10.1109/55.974579.

[2] Dominique Mangelinck, Khalid Hoummada, Alain Portavoce, Carine Perrin, Rachid Daineche, Marion Descoins, David J. Larson, Peter H. Clifton, Three-dimensional composition mapping of NiSi phase distribution and Pt diffusion via grain boundaries in Ni2Si, Scripta Materialia, 62 (2010) 568-571; DOI: 10.1016/j.scriptamat.2009.12.044.

[3] J. Demeulemeester, D. Smeets, C. Van Bockstael, C. Detavernier, C.M. Comrie, N.P. Barradas, A. Vieira, A. Vantomme, Pt redistribution during Ni(Pt) silicide formation, Applied Physics Letters, 93 (2008) 261912; DOI: 10.1063/1.3058719.

[4] N.P. Barradas, A. Vieira, Artificial neural network algorithm for analysis of Rutherford backscattering data, Physical Review E, 62 (2000) 5818; DOI: 10.1103/PhysRevE.62.5818.

[5] J. Demeulemeester, D. Smeets, N.P. Barradas, A. Vieira, C.M. Comrie, K. Temst, A. Vantomme, Artificial neural networks for instantaneous analysis of real-time Rutherford backscattering spectra, Nuclear Instruments & Methods B, 268 (2010) 1676-1681; DOI: 10.1016/j.nimb.2010.02.127.

[6] F. Corni, B. Grignaffini Gregorio, G. Queirolo, J.P. Follegot, Dilute NiPt alloy interactions with Silicon, Applied Surface Science, 73 (1993) 197-202; DOI: 10.1016/0169-4332(93)90166-9.

[7] J. Demeulemeester, D. Smeets, C.M. Comrie, N.P. Barradas, A. Vieira, C. Van Bockstael, C. Detavernier, K. Temst, A. Vantomme, On the growth kinetics of Ni(Pt) silicide thin films, Journal of Applied Physics, 113 (2013) 163504; DOI: 10.1063/1.4802738.

[8] J. Demeulemeester, D. Smeets, C.M. Comrie, C. Van Bockstael, W. Knaepen, C. Detavernier, K. Temst, A. Vantomme, The influence of Pt redistribution on Ni1-xPtxSi, Journal of Applied Physics, 108 (2010) 043505; DOI: 10.1063/1.3455873.

[9] C.C. Theron, J.A. Mars, C.L. Churms. J. Farmer, R. Pretorius, In situ real-time RBS measurements of solid state reaction in thin films, Nuclear Instruments & Methods B, 139 (1998) 213-218; DOI: 10.1016/S0168-583X(97)00946-4.

[10] J. Rennie, S.R. Elliott, C. Jeynes, Rutherford backscattering study of the photodissolution of Ag in amorphous GeSe2, Applied Physics Letters, 48 (1986) 1430-1432; DOI: 10.1063/1.96879.

[11] R.S. Averback, L.E. Rehn, W. Wagner, H. Wiedersich, P.R. Okamoto, Kinetics of radiation induced segregation in Ni-12.7 at.% Si, Physical Review B, 28 (1983) 3100-3109; DOI: 10.1103/PhysRevB.28.3100.

[12] N.P. Barradas, C. Jeynes, Advanced physics and algorithms in the IBA DataFurnace, Nuclear Instruments & Methods B, 266 (2008) 1875-1879; DOI: 10.1016/j.nimb.2007.10.044.

[13] C. Jeynes, N.P. Barradas, E. Szilágyi, Accurate determination of quantity of material in thin films by Rutherford backscattering spectrometry, Analytical Chemistry, 84 (2012) 6061-6069; DOI: 10.1021/ac300904c.

[14] C. Jeynes, N.P. Barradas, P.K. Marriott, G. Boudreault, M. Jenkin, E. Wendler, R.P. Webb, Elemental thin film depth profiles by ion beam analysis using simulated annealing – a new tool, Journal of Physics D: Applied Physics, 36 (2003) R97-R126; DOI: 10.1088/0022-3727/36/7/201.

[15] N.P. Barradas, A. Vieira, P. Patrício, RBS without humans, Nuclear Instruments & Methods B, 190 (2002) 231-236; DOI: 10.1016/S0168-583X(01)01249-6.

[16] N.P. Barradas, A. Vieira, R. Patrício, Artificial neural networks for automation of Rutherford backscattering spectroscopy experiments and data analysis, Physical Review E, 65 (2002) 066703; DOI: 10.1103/PhysRevE.65.066703.

[17] A.M. Müller, A. Cassimi, M. Döbeli, M. Mallepell, I. Monnet, M.J. Simon, M. Suter, H.-A. Synal, A new mini gas ionization chamber for IBA applications, Nuclear Instruments & Methods B, 269 (2011) 3037-3040; DOI: 10.1016/j.nimb.2011.04.078.