Introduction

Metallic contamination is a serious threat in the semiconductor industry. Throughout the manufacture of integrated circuits, metallic contamination can occur at any step brought by the tools used in the production and/or human activity. It can be accidentally introduced during the formation of silicon ingots or by the different fabrication processes as photolithography, annealing or cleaning among other steps [1]. The presence of contamination on the surface brings numerous impacts, for example the deformation of the surface state or change of the surface properties. High temperatures often used during the fabrication processes induce the diffusion of metallic contaminants into the bulk. As a result, integrated circuit encounters electrical issues, such as an excessive leakage current or decrease of carriers lifetime leading to the degradation of the device performance [2]. Therefore, the earlier the contaminants are detected on the production line, the better. That is why cleanrooms must keep the level of contamination below a specified threshold by performing regular controls of manufacturing tools and the environment, which can differ from one manufacturing facility to another.

Currently, the most widely used method in the industry of microelectronics for metallic contaminants surface analysis are Total X-Ray Fluorescence (TXRF) and Vapor Phase Decomposition—Inductively Coupled Plasma Mass Spectrometry (VPD-ICPMS). Both methods are well mastered on smooth silicon wafers but are confronted to some limitations regarding rough and patterned wafers. Indeed, during VPD-ICPMS, the collection step uses a chemical droplet held by a nozzle. The presence of roughness can retain the droplet leading to loss of contaminants [3]. For TXRF, it has been observed with Angle-Dependent TXRF that roughness can lead to interferences between the incident and reflected rays. This results in distortion of the intensity curves and the increase of the background noise, giving unrepresentative and unreliable results [4]. Processed wafers with surface roughness induced by mechanical polishing, chemical polishing, or wet cleaning are difficult to analyse. Thus, at the present time, the monitoring of metallic contamination is mainly performed on smooth wafers. This practice considers that if contamination is not found on the smooth wafers, then it is not present on the processed ones. The monitoring of processed wafers becomes necessary and as TXRF is a non-destructive method, it is a method of choice compared to VPD-ICPMS.

The aim of this study is to identify the influence of surface roughness on TXRF parameters and spectra and to define the value of the roughness from which the TXRF measurements are not reliable anymore. Wafers with different level of roughness were studied, generated by various processes commonly used in the microelectronic fabrication line. Then, the impact of roughness on the detection and quantification of metallic elements was highlighted by analysing intentionally contaminated rough wafers with selected elements at various concentrations.

Total X-ray fluorescence

TXRF allows non-destructive and fast analyses. Based on electron-matter interaction, it uses X-Ray beams with specific angles and energies allowing total reflection with low penetration in the wafer (Table 1). TXRF300 equipment from Rigaku uses three monochromatic beams generated by a rotating tungsten anode. Limits of detection for this method are around 1 × 1010 at/cm2, depending on the element. Moreover, it provides a mapping for the localisation of contaminants, which can lead to the origins of metallic contamination [5].

TABLE 1 Energy and incident angle of the three monochromatic beams used for the study.
Full size table

Analyses are usually carried out in sweeping mode which consists in the multiplication of measurement points with a short acquisition time while having a maximum surface coverage. For example, a mapping in 93 points spaced of 17 mm for 200 mm wafers and 133 points spaced of 22 mm for 300 mm wafers allows a surface coverage between 70 and 90% as illustrated by Fig. 1. Each point covers a surface of approximatively 4 cm2, distributed in a concentric shape with an edge exclusion of 15 mm for beam 1 and 2, and 25 mm for beam 3. The acquisition time per point for both wafer sizes is 5 s for each beam. Those are the parameters used for this study.

Figure 1
figure 1

Distribution of measurement points by TXRF on (a) 200 mm and (b) 300 mm wafers.

Full size image

Results and discussion

Impact of roughness on TXRF parameters

Several parameters were checked to ensure a correct analysis:

  • Number of counts per second (cps) of Silicon (Si), coming from the analysed Si wafer substrate,

  • Number of counts per second (cps) of Tungsten (W), coming from the anode of the TXRF,

  • Dead rate (%), representing the part of signal that cannot be analysed by the detector because of its saturation by the number of incoming photons,

  • Background noise in counts per second (cps) coming from the source of X-Rays and from the interaction with the substrate. With greater interactions, a higher fraction of Si coming from the substrate and W from the incident beams contribute to the background noise, consequently it leads to wide Gaussian like peaks instead of sharp ones and causes interferences with peaks of nearby elements [6].

Those parameters must be in a range of values to consider analyses reliable, and they can differ from one tool to another. Ranges are given by the equipment supplier. Since metallic contamination is always carried out on smooth Si wafers, numbers of cps in Si and W, and percentage of dead rate are usually in the ranges listed in the last line of Table 2. Those ranges are applicable for the TXRF used in this study with the experimental parameters mentioned in "Total X-ray fluorescence" Section. Generally, the variation of the parameters inside the acceptable ranges can be due to the grade of the silicon wafers, the crystallography, the quality and the thickness of the oxide.

TABLE 2 TXRF parameters of W and/or Si cps and dead rate obtained for the samples with various roughness, the acceptable range of these parameters are listed on the last line.
Full size table

Table 2 presents the results of eight samples with various roughness, they are either reference wafers or wafers produced by grinding, wet cleaning or chemical mechanical polishing. The material preparation is detailed in "Wafers fabrication" Section and roughness measurement in "Roughness measurement" Section.

For all three beams, parameters stay in the acceptable range for a reliable analysis until roughness reaches around 3 nm. With the increase of roughness, the values of the parameters also increase, except for the dead rate on beam 1 which stays at 20% whatever the roughness. The increase of the parameters is all the more striking for samples 6 and 7 (grinded wafers rough of 9 and 12 nm). For sample 8, the non-polished back-side of 200 mm Si wafer with the highest roughness at about 192 nm, shows the highest dead rate. All the parameters are above the acceptable ranges, except Si cps on beam 2. The roughness topography of this wafer is different from the other wafers in terms of uniformity due to the process, it might influence the interaction of the incident beam and the substrate.

Moreover, it seems that the TXRF parameters for beams 2 and 3 are more impacted by roughness than beam 1 parameters. It may be due to the incident angle used for each beam. Indeed, beam 1 angle is at 0.5°, higher than the angle of beam 2 and beam 3 which are, respectively, at 0.08° and 0.05°, more grazing angles. Consequently, beams 2 and 3 are more impacted by the surface roughness of wafers.

The impact of background noise and its consequences are visible on the spectra. In Fig. 2, the spectra comparison of sample 2 (200 mm Si reference), sample 6 (9 nm grinded wafer) and sample 8 (non-polished back-side 200 mm) shows for beam 1 a faint widening of W-Mα peak (main peak on the right of the spectrum). In fact, W-Mα (1.78 keV) and Si-Kα (1.74 keV) peaks overlap, the combination of both can lead to an even wider peak which depends on their number of counts per second. The higher the count, the wider the peak. It covers the nearest element Al at nearly 1.45 keV and can also have an impact on Mg at nearly 1.25 keV.

Figure 2
figure 2

TXRF spectra of beam 1, (a) Si 200 mm reference front-side, (b) grinded wafer RMS 9 nm and (c) Si 200 mm reference back-side.

Full size image

On beam 2 spectra (Fig. 3), the widening of Si-Kα (1.74 keV) and W-Lβ (9.67 keV) is also observed, respectively, positioned on the left and right side of the spectrum. Moreover, with important Si and W cps, sum peak belonging to Si-Kα (3.48 keV) and escape peak belonging to W-Lβ (7.93 keV) are present. Altogether, these peaks can interfere with several elements: K, Ca, Cu or Zn. A rise of background noise is also discerned which accentuates the widening of Si and W peaks. Moreover, several elements are present, Cl and Ca coming from the handling of wafers and Fe brought by the grinding process. Zn is also found but the W-Lβ peak interferes with it.

Figure 3
figure 3

TXRF spectra of beam 2, (a) Si 200 mm reference front-side, (b) grinded wafer RMS 9 nm and (c) Si 200 mm reference back-side.

Full size image

Background noise on beam 3 spectra (Fig. 4) is the main parameter to change greatly in the presence of roughness, elements can be drowned in the background like Ga, Ta and Mo peaks on sample 8 (Si 200 mm back-side). A widening of Si peak is also observed (main peak on the left side of the spectrum).

Figure 4
figure 4

TXRF spectra of beam 3, (a) Si 200 mm reference front-side, (b) grinded wafer RMS 9 nm and (c) Si 200 mm reference back-side.

Full size image

Impact of roughness on metallic contamination quantification

The acceptable ranges of TXRF parameters for reliable analysis were defined on smooth silicon. As studied above, roughness increases those parameters and induces higher limits of detection. Nevertheless, quantification is not impossible, so to understand the impact of roughness on metallic contamination quantification, rough wafers with values of Si and W intensity out of the acceptable ranges were selected: grinded wafers of 12 nm RMS roughness. Intentional contamination was performed by spin-coating at concentrations of 1 × 1010, 1 × 1011 and 1 × 1012 at/cm2, one sample per concentration. Several elements covering the energy range of the three beams were picked ("Intentional contamination" Section). Wafers have not been cleaned before intentional contamination to avoid the modification of the surface condition.

One grinded wafer without intentional contamination was taken as a reference. As seen in Table 3, W and/or Si cps are above the acceptable ranges. Beam 1 parameters are in the same ranges for the 4 wafers. On beam 2 and 3 parameters values are smaller on the contaminated wafers. A fluctuation from TXRF could be at the origin of the difference between the not contaminated and the contaminated rough wafers.

TABLE 3 TXRF parameters measured before and after intentional contamination on rough wafers obtained by grinding.
Full size table

Diffraction generated by roughness increases the background noise, which is related to the Low Limit of Detection (LLD) [7]:

$${\text{LLD}} = 3 \times \frac{{\sqrt {{\text{BG}} \times T} }}{T} \times \frac{C}{S},$$
(1)

where BG is the background in cps, T the measurement time in s, C the concentration of the analysed element in at/cm2 and S the intensity of the analysed element in cps.

Table 4 presents the LLDs for the reference Si wafer (sample 2) and the not contaminated grinded wafer (RMS = 12 nm). The rough wafer has on average LLDs two times higher than those of the reference for each element due to the rise of the background.

TABLE 4 TXRF LLDs on bare Si wafer and not contaminated rough wafers.
Full size table

These LLDs explain the non-quantified elements on the intentionally contaminated wafer at 1 × 1010 at/cm2. Table 5 presents the quantification of the elements of the not contaminated reference wafer (A) and the intentionally contaminated wafers at 1 × 1010 (B), 1 × 1011 (C) and 1 × 1012 at/cm2 (D).

TABLE 5 Quantified concentrations by TXRF on the non-contaminated 12 nm RMS rough wafer (sample A). Quantified concentrations by TXRF and VPD-ICPMS of the 12 nm RMS rough wafers contaminated at 1 × 1010 (sample B), 1 × 1011 (sample C) and 1 × 1012 at/cm2 (sample D), and LLDs of both methods are given.
Full size table

First, because of the absence of preliminary cleaning, Na, Fe and Zn contamination were observed on the not intentionally contaminated wafer, probably coming from the handling and the grinding process. It has to take into account that the intentional contamination protocol is adapted for smooth wafers, the amount of contaminants might differ on rough surface and from one roughness to another since a rougher wafer has more contact surface with contaminants. But the elements used for intentional contamination do not change the surface roughness, so the variation of quantification cannot be attributed to surface roughness. Furthermore, for the concentration quantification by TXRF and VPD-ICPMS, the experimental error must be taken into account (operator and equipment itself). On smooth wafers, the concentration obtained by TXRF can variate around ± 5% on beam 1, ± 20% on beam 2 and ± 30% on beam 3. For VPD-ICPMS analyses, the variation is ± 15%.

At a target contamination of around 1 × 1010 at/cm2 (sample B), all the concentrations quantified by VPD-ICPMS are below the TXRF LLDs, which explains why they are not detected and quantified by TXRF. Except for Fe and Zn, VPD-ICPMS gives concentrations above the TXRF LLDs. However, Zn peak (at 8.646 keV) is interfered by W-Lβ peak (at 9.67 keV) and is not quantified. For Fe, TXRF gives a concentration of 6.3 × 1010 at/cm2, which corresponds to the concentration found by VPD-ICPMS at 4.5 × 1010 at/cm2. Fe can be well quantified at this roughness and this level of contamination.

For the wafer with target contamination around 1 × 1011 at/cm2 (sample C), almost all elements measured with beam 2 were detected, except Zn (at 8.64 keV) since its peak is interfered by the W-Lβ peak. To be noted that Cu is not well collected by VPD-ICPMS, only 40 to 60% of contaminants are collected on smooth Si wafers and it could be even lower with surface roughness. On beam 3, the Mo peak is not identified by the TXRF software because of the background noise which is almost as high as the peak. Then, Ga is not quantified, the concentration given by VPD-ICPMS shows that it is below the TXRF LLD. For beam 1, no elements are detected as Al is interfered by W-Mα, and Na and Mg are near the TXRF LLDs.

At 1 × 1012 at/cm2 (sample D), all the elements were quantified correctly as the concentrations given by TXRF correspond to the ones obtained by VPD-ICPMS. The only exception is for Al, which is greatly limited by the widening of the W-Mα peak. Zn does not seem to be interfered by W-Lβ peak, TXRF and VPD-ICPMS analysis give the same concentration. For Ga (at 9.25 keV), its peak overlaps with Ta’s secondary peak (Ta-Lβ1 at 9.34 keV), Ta-Lβ1 contributes to Ga’s peak. Therefore, Ga is overestimated.

Thus, for detection at 1 × 1010 at/cm2, TXRF LLDs are too high and do not allow the quantification of the elements on a wafer with roughness RMS superior or equal to 12 nm. From 1 × 1011 at/cm2, most elements are quantified correctly. Only Na, Mg, Al (element of beam 1) and Zn are mainly affected by the interference with the W peaks, whereas Ga and Mo are drowned in the background noise. At 1 × 1012 at/cm2, all the elements are quantified correctly, except Al.

Conclusions and perspectives

In this paper, the influence of the surface roughness on the TXRF measurements was studied. Roughness is known to be a key parameter in the reliability of TXRF results, measured samples are usually smooth to prevent diffraction. Hence, there is not much information on the limit values of roughness allowing a reliable measure and on the generated effects if a sample is rough. Several wafers with different surface roughness values were prepared using such processes as CMP, wet chemical etching and grinding. All wafers were characterized at first by AFM and/or optical profilometer to assess the roughness values. Afterwards, TXRF sweeping measurement showed that for a RMS roughness above 3 nm, different standard parameters of the TXRF increase and go out of the acceptable ranges. This is the case for the count per second of Silicon and Tungsten, as well as the dead rate percentage. Consequences on spectrum are an increase of the background noise which leads to an increase of LLDs and the enlargement of Si and W peaks that interfere with nearby elements, especially Al. As roughness increases, those effects affect more and more the analysis. The values of TXRF LLDs on a smooth Si wafer and on a grinded wafer with a RMS of 12 nm, were compared. The LLDs is two times higher on the grinded wafer, it can influence the detection and the quantification of elements.

To understand the influence of roughness on metallic contamination detections and quantifications, grinded wafers with a RMS roughness of 12 nm were intentionally contaminated. At a contamination of 1 × 1010 at/cm2, detection and quantification are almost impossible because of the increase of the background noise leading to higher LLDs. Most elements become quantifiable at 1 × 1011 at/cm2, except for peaks near W and Si peaks which are interfered, like Na, Mg, Al and Zn. At 1 × 1012 at/cm2, all elements are quantified correctly except Al, this element is the nearest from the W peak and is constantly affected by its widening. VPD-ICPMS on the rough wafers indicated whether the concentration of each element was below the TXRF LLDs or not. Which helped in explaining the non-detected elements by TXRF.

TXRF technique is sensitive to roughness due to the specificity of the tool. With incident angles less than 1 degree, many phenomena are observed making it difficult to detect and quantify metallic contaminants. However, if the concentration of elements exceeds 1 × 1011 at/cm2, it is possible to detect and quantify correctly. The cleanroom is ruled by different metallic contamination categories: very low metallic contamination for front-end of line tools (< 5 × 1010 at/cm2), intermediate range for middle end of line tools and higher contamination for back end of line tools (> 1012 at/cm2). For the latter case, quantification on rough wafers is now accessible.

To go further, methods less sensitive to surface roughness like Vapor Phase Decomposition or Liquid Phase Decomposition followed by Inductively Coupled Plasma Mass Spectrometer could be studied for rough and patterned contaminated wafers. These involve the will to sacrifice the wafers as those methods are destructive.

Materials and methods

Wafers fabrication

Silicon wafers (200 mm or 300 mm, p-doped, oriented < 100 >) with different surface roughness were prepared using various processes from the semiconductor industry: grinding, Chemical Mechanical Polishing (CMP) and wet cleaning.

The grinding process is generally used to thin and polish front and/or back sides of wafers with wheels made of abrasive grains. Depending on the size, the sharpness of the grains, the hardness, the porosity and quality of the bond as well as the polishing time, different ranges of roughness can be obtained [8]. Here, roughness was produced by in-feed single side grinding tool (DISCO Corporation) in two steps. First a coarse grinding at 40 µm/min removes most of the thickness required and then a fine grinding at 10 µm/min reduces the roughness. Two 200 mm Si wafers were prepared using a coarse grinding removing a thickness of 30 µm. Then two levels of roughness were generated with the fine grinding step by removing two different thicknesses, 5 µm and 25 µm, respectively.

CMP is a soft polishing process by chemical and mechanical means. It is used to flatten a deposit or remove impurities introduced during previous processes [9]. In this study, a 300 mm Si wafer was processed by grinding and then a CMP step of 5 s was applied to decrease the roughness.

Wet cleaning is a chemical procedure used to clean wafers. An etching step can modify the surface texture [10]. In this study, two 200 mm Si wafers were etched, respectively, during 30 and 60 s with HF/HNO3 solution.

Roughness measurement

The parameter used in this study for comparing the roughness of wafers is the Root Mean Square (RMS), the quadratic distribution of roughness on the analysed surface [11]. It was measured by Atomic Force Microscopy (AFM) with a Bruker Fastscan in TappingMode™ and a 5 nm nominal radius silicon tip. For each wafer, scans of 1 × 1 µm and 10 × 10 µm were acquired. Raw data were flattened by a line-by-line bow removal (polynomial fit, order 1) [12]. Line-by-line leveling is the most used and the simplest leveling method to remove the curvature or the slope background. As most of the measured samples have small roughness, a curvature or a slope could misrepresent the real roughness. For the roughest wafers, an optical profilometer Contour GTX from Bruker was used in vertical scanning mode. Each wafer was measured on three locations: middle, half-radius and 1 mm from the edge. Final RMS roughness is an average of the three measurements.

Table 6 presents the roughness measurements of Si wafers after different processes. Non-processed Si wafers are used as a reference, in 200 mm and 300 mm. They have the lowest roughness with a RMS of 0.14 nm and 0.13 nm, respectively. The wafer processed by CMP has a roughness slightly higher with a RMS of 0.19 nm. With more aggressive processes using chemicals and grinding, the surface roughness increases up to 12 nm. Finally, non-polished back-side of 200 mm Si wafer has the highest roughness, about 192 nm.

TABLE 6 Roughness of wafers (200 or 300 mm diameter) produced by different microelectronic processes measured by AFM or optical profilometer.
Full size table

Intentional contamination

For intentional contamination, three different concentrations: 1 × 1010, 1 × 1011 and 1 × 1012 at/cm2 were performed by spin-coating (POLOS 300). The following elements were chosen to cover all spectral range of the three TXRF beams: Al, Mg, Na, Ti, Fe, Ni, Cu, Zn, Ga, Mo and Ta. Single element solutions (1000 µg/mL in HNO3, 5%) were provided by Merck, Alfa Aesar, Chem Lab, Czech Metrology Institute, VWR International, Spex or Peak performance. Each element was added to a solution of HNO3, 1%. The wafer was placed in the spin-coater, a volume of the contaminated solution was poured on the wafer to cover the whole surface and then let rested for 1 min of exposure time. Then, the wafer was rotated at 2000 rpm (rotation per minute). This action allows a homogeneous contamination and the drying of the wafer.

VPD-ICPMS analysis (VPD300 from Rigaku and ICPMS 8800 from Agilent) was performed on the intentionally contaminated wafers to ensure that contamination was deposited at wanted concentrations. Collection efficiency representing the ratio of collected concentration to the initial concentration of contaminants is above 95% for all elements except for Cu between 40 and 60% [13] on smooth silicon wafers. Also, limitation of VPD-ICPMS on rough surfaces mentioned above was not observed on the wafers used in the study discussed in "Impact of roughness on metallic contamination quantification" Section .