Mixture Analysis without Separation Techniques

Mixture Analysis without Separation Techniques

Facilitating impurity detection with multivariate curve resolution

WaferTech, a U.S. semiconductor foundry, produces silicon wafers for a wide variety of integrated circuits for consumers throughout the automotive and medical electronic sectors. Achieving a high-quality production process for the myriad product-specific recipes that the foundry must create requires extreme attention to

click to enlarge Figure 1: ATR-IR spectra of fumed silica powder lots

performance controls for robots, sensors, temperature/pressure/flow controllers, environmental properties, human operators and process materials. One cannot fully predict product quality until the process sequence is completed. However, at the same time, one cannot simply roll the dice in the face of an uncertain process condition when more than 10 weeks of expense and labor are at risk.

Verification of raw materials, including silicon wafers, photoresists, etching solvents and gases, and silicon oxide-based insulators, is accomplished by supplier competency audits, quality assurance (QA) analytical reports (certificates of analysis), and in-house chemical analysis. In this case, infrequent chemical sampling and analysis is an accepted practice because of excellent raw material quality control. Evaluation of raw materials at the point of use (POU) is of greater importance because handling, delivery and process influence can alter predicted performance.

POU sampling and analysis is the cornerstone for verification of process material integrity; and analytical capabilities such as inductively coupled plasma mass spectroscopy (ICP-MS), graphite furnace atomic absorption (GFAA), ion chromatography, titration, bacteria monitoring, and fourier transform infrared spectroscopy (FTIR) are routinely employed to verify material assay and impurity content. Rapid turnaround time is critical to minimize response time when material property control limits are exceeded.

Analytical capabilities also are applied to process excursion control activities in order to narrow the scope of investigation for determining the primary failure mode. Such non-routine analytical exercises require instrument competency, familiarity with both sample and instrument responses, and the ability to essentially pull a needle from a haystack and to verify that it is, indeed, a needle. Note that the analytical focus of semiconductor process prioritizes metal impurities and particles because of the intrinsic impact upon electrical performance and circuit disruption potential. Organic impurity characterization thus tends to be a pseudo blind spot, because typical techniques show an apparent lack of contrast between constituents that can profoundly impact the ability to interpret analytical subtleties. Most material manufacturing processes involve mixtures, whether these are part of the actual product or related support materials. When an excess of costly material is used, when poor-quality consumable materials lengthen processing time or cause extra processing steps to be required or when entire lots fail due to the presence of unacceptable impurities, quick troubleshooting and resolution of these issues can become business-critical. In this context, scientists are often called upon to use instrumental analysis to rapidly analyze mixture content that, when out-of-control, can result in product rejection at a potential cost on the order of millions of dollars.

Analytical challenges

Spectral interpretation of traditional molecular spectroscopy techniques can be especially challenging when a sample is a mixture of substances. The analysis of complex samples generally calls for prior extraction and separation of isolated sample components — if feasible — in conjunction with the application of molecular spectroscopy techniques to identify and semi-quantify each chemical entity found in the mixture. However, for solid-state materials, inorganic mixtures, highly reactive substances or organic reaction intermediates, there is often no physical or chemical means of isolating components.

Where organic or dielectric material properties must be characterized, FTIR molecular spectroscopy is generally used for rapid and non-destructive analysis. Spectral data is readily obtained, but spectral interpretation of sample mixtures to determine assay or trace impurity identity for chemically reactive materials exceeds the skill set of most laboratory personnel. The rich spectral profile with limited resolution of key species precludes direct, meaningful

click to enlarge Figure 2: Spectra for chemical components detected in raw silica powder lots

interpretation. Sophisticated chemometric algorithms are commercially available for turn-key applications, such as oxide dopant concentration, but adaptation to a real-world non-routine chemical matrix is unrealistic in light of the potential instability of the matrix. Without extended time to create a calibration set for a process excursion event, raw spectral interpretation is limited at best.

Assays can be determined by slower and preparation-intensive chromatographic techniques and titrations, as well as by spectral quantitation using FTIR and/or other molecular spectroscopies. The advantage of preparative methods is that impurities can be characterized and interpreted more easily once extracted and isolated. The disadvantage is that using such techniques risks compromising the sample and its relevance with respect to the original sample matrix. In contrast, non-destructive molecular spectroscopy techniques applied in situ, with little to no sample pretreatment, make for very quick measurement but require longer interpretation times due to significant data interpretation challenges. Increasingly, multivariate curve resolution is being used as a means for digitally separating pure components, thus helping to quickly overcome the interpretation barrier for mixture interpretation by molecular spectroscopy techniques.

Experimental algorithm test

WaferTec’s engineers selected process troubleshooting applications involving POU organic etching bath solutions and abrasive powders (to be blended into slurry) in order to test the capabilities of a hybrid FTIR-diamond attenuated total reflection (ATR) system with an experimental interpretation algorithm. Each application challenged conventional analysis methods because of limited contrast of baseline versus out-of-control samples. The diamond ATR sampling accessory was chosen because of its chemical inertness and durability that suppress sample history and sample sequence bias. The Simple Interactive Self-modeling Mixture Analysis (SIMPLISMA) algorithm was selected in order to evaluate its ability to isolate independent spectral responses with no prior familiarity with sample composition.

The ability to isolate and identify components using multivariate curve resolution is gaining favor due to the introduction of chemometric tools easily applied by experts and non-experts alike. Tools like the SIMPLISMA experimental interpretation algorithm are technically applicable to any set of one-dimensional spectra where:
&#149 constituent concentrations exceed three times the minimum detect limit
&#149 distinguishing constituent features are present
&#149 constituents exhibit concentration changes across the data set

Mathematically, the curve resolution equation can be expressed essentially as D = C × P, where D is the original spectral matrix, C is the relative concentration matrix, and P is the pure component spectra matrix. SIMPLISMA works by finding pure variables to resolve these complex spectral matrices. In this approach, the intensities taken at a pure variable for each component are used to provide a relative estimate for the concentration matrix C. A pure variable is any x variable where only one of the mixture components shows a significant response &#151 all other components remaining essentially silent. Using the pure component intensities for C, the pure component spectra matrix P can be resolved by least squares regression.

In order to find pure component variables, SIMPLISMA calculates a purity spectrum. This is essentially a modified expression of the relative standard deviation as follows

where p_jn is the purity variable value; s_j and µ_j are the standard deviation and mean of the j^th variable, respectively; and α is a constant offset designed to prevent artificially high purity values in the cases where the mean value approaches zero. This equation is referred to as a purity function. It was shown that the higher the purity variable value, the higher its relative component purity for this variable.¹ Therefore, if we plot the purity value p_j at each variable j this produces the so-called ‘purity spectrum,’ and the maximum value in the plot will correspond to the first ‘pure variable’ component. To avoid picking the same component twice, the effect of the first pure variable is mathematically eliminated from the purity spectrum. SIMPLISMA resolves the components one-by-one. For more accurate descriptions and details on how the algorithm works, please refer to the literature.^1,3–5

Detecting trace components

It is particularly difficult to detect and identify trace impurities in mixtures with a strongly absorbing matrix background. To illustrate this, let us look at the infrared (IR) spectra of raw silica powder that were measured neat by diamond ATR. As seen in Figures 1a and 1b, simply comparing the IR spectra does not reveal any obvious difference. A comparison across 52 lots of raw

click to enlarge Figure 3: ATR-IR spectra of etch bath solution

powder in Figure 1c shows some differences, but it is difficult to interpret which features are due to sample preparation and which are due to genuine lot differences. The obvious baseline shift is primarily attributable to inconsistencies in the contact and coverage of powder sample against the ATR crystal. Despite this apparent lack of differences, certain lots were found to fail. At this point, WaferTec was able to look for clues using SIMPLISMA.

Raw silica powder is used in the formulation of a chemical mechanical polishing (CMP) slurry used to polish silicon wafers. Certain CMP lots were observed to have poorer polishing performance. Bad lots appeared to aggregate and agglomerate aggressively in water over time, causing performance instability. Unfortunately, once blended in a several-thousand-gallon tank, it is too late to return the product. Therefore, the root cause of the poor performance had to be characterized by indirect means to restore abrasion process control.After performing a group baseline correction to correct for the baseline shift, the data set was treated with SIMPLISMA. Three components were identified and extracted with pure variables maxima at 1060 cm,^-1 1215 cm^-1 and 1742 cm,^-1 respectively. The residual standard deviation does not yield significant improvement beyond three components, and trying to extract a fourth component yields negative artifact peaks and baseline curvature. Figure 2 shows the spectra of the three extracted main components.

The primary component, shown in blue, was matched to stable high-density raw silica. The next component, in green, corresponds to reactive low-density raw silica. The third component, in red, shows bands that correlate to an organic impurity. Good lots had more of the high density silica and less of the low density component, with negligible amounts of organic impurity. Bad lots had elevated low density raw silica and/or organic impurity content.

In this example, we see that the SIMPLISMA algorithm was able to extract and distinguish between two different crystal forms and an impurity component without having to run standards or rely on a priori knowledge. This type of analysis took seconds to perform and helped produce a quick and reliable batch rejection criterion.

Kinetic profile of a batch process

There are many instances where knowing the overall kinetic profile of a mixture would be useful, but where this is not readily available. For example, in order to estimate how long an etch bath retains its effectiveness, WaferTec’s engineers needed to extract kinetic profiles of active constituents. It can be extremely challenging to extract kinetic information by classical methods for blended materials with dynamic composition or with spectral interactions by blend constituents. For example, while classical analysis methods are able to predict water content, they provide little indication of the overall kinetics with respect to the active etch ingredient within the complex ATR-IR time study shown in Figure 3.

After constant background correction, masking of the diamond artifact, and treatment with SIMPLISMA, pure component spectral and concentration profiles for water, the hydroxyl amine active ingredient and alcohol amine can be obtained as shown in Figure 4. A net hydroxyl amine accumulation over the bath life suggests that the depletion of the etchant is more than counterbalanced by the concentration effect due to water evaporation.

This experiment validated the extension of the useful life of the bath by a factor of three, thus reducing costs. The kinetic interpretation achieved through random batch sampling did not require a formal kinetic experiment design, and is representative of long-term variance and stability.

Advantages

The SIMPLISMA algorithm is particularly useful for mixture deformulation where the chemistry or spectroscopy of a sample is unknown (or only partially known). This is because it does not require a priori knowledge of the sample system. Standards and reference material are not needed to run SIMPLISMA.

The algorithm produces spectroscopic plots and results that can be responsibly and interactively evaluated by using spectroscopic judgment (without requiring deep chemometric or mathematical expertise). Its interactive nature

click to enlarge Figure 4: Spectral profile for three primary components extracted with SIMPLISMA (top) and corresponding relative concentration profile trends for water and hydroxyl amine in relation to the alcohol amine (bottom)

is practical when troubleshooting in industrial environments, since it is difficult in such settings to ensure controlled experiments, obtain replicates, or avoid uncontaminated samples. The instant feedback provided by the spectroscopic plots extracted for each component means that the scientist can recognize when things go wrong as the results unfold. As a result, it is possible to direct the procedure by using chemical knowledge of expected components. This means that corrective actions can be taken immediately and makes it less likely that artifacts will be misinterpreted as real features. The results obtained from SIMPLISMA also can provide valuable guidance in preparing and interpreting more complex chemometric models.

Discussion

In a sense, the SIMPLISMA algorithm bridges the gap between process and analytical chemists. On the one hand, process engineers generally have a good understanding of expected sample composition, without necessarily understanding the underlying spectroscopy. On the other hand, spectroscopists can properly interpret spectral changes but do not always have insight into how process may affect the chemistry. It is to be noted that the algorithm provided insight on reaction or aging kinetics in a fairly easy manner as compared to more classical methods.

The combination of FTIR with diamond ATR sampling and SIMPLISMA interpretation successfully demonstrated to WaferTec’s engineers an ability to derive sample assay and the presence of trace impurities from a series of randomly selected process samples with no externally derived sample condition knowledge. There also are opportunities to evolve SIMPLISMA into a learning algorithm that treats each new data set sample as an unknown (to characterize its distinction) prior to addition to the hierarchal dataset. For priority process excursion support, samples from relatively high risk material systems can be included in the data set to add baseline characteristics, helping to uncover unknown sample process conditions.

We have shown two complex matrix applications where SIMPLISMA was able to detect and resolve components in commercial mixtures using analytically sensitive ATR-IR spectra. The results allowed the variance to be characterized and identified, which is the first step toward control and elimination. As such, the algorithm can be considered a valuable tool for reproducibility and repeatability studies, finding batch-to-batch raw material variations, identifying product contaminations, and performing comparative analysis of good versus bad product batches.

References

1. Windig, W.; Guilment, Journal of Analytical Chemistry 63, 1425–1432 (1991).
2. ACD/UV-IR Manager, v10, Advanced Chemistry Development, Toronto, Canada, www.acdlabs.com/uvir, 2007.
3. Windig, W.; Stephenson, D.A., Journal of Analytical Chemistry 64, 2735–2742 (1992).
4. Windig, W., Chemometrics and Intelligent Laboratory Systems 36, 3–16 (1997).
5. Bogomolov, A.; Hachey, M.; Williams, A., “Software for Interactive Curve Resolution Using SIMPLISMA,” in progress in Chemometric Research (editor Alexey L. Pomerantsev), Nova Science Publishers, New York, Chapter 10,199–135 (2005).

Stephen Hill is a supplier quality management engineer at WaferTech. Michael Boruta is an optical spectroscopy product manager, and Michel Hachey is a technical specialist at Advanced Chemistry Development. They may be contacted at [email protected].

Acronyms
ATR Attenuated Total Reflection | CMP Chemical Mechanical Polishing | FTIR Fourier Transform Infrared Spectroscopy | GFAA Graphite Furnace Atomic Absorption | ICP-MS Inductively Coupled Plasma Mass Spectroscopy | IR Infrared | POU Point of Use | QA Quality Assurance | SIMPLISMA Simple Interactive Self-modeling Mixture Analysis