Estimating bauxite quality USING FTIR Spectroscopy and multivariate calibration

Sharon L Eyer

Research and Development Department, Alcoa of Australia Ltd

PO Box 161, Kwinana, Western Australia, 6167

ABSTRACT

Consistency of bauxite supply has a significant effect on alumina quality. Excursions in alumina content and impurity levels can affect aspects such as circuit A/TC and formation of desilication product potentially causing problems with process control.

FTIR spectroscopy can be a useful tool for estimating the quality of an ore body based on exploration samples. Furthermore, it can highlight regions of bauxite that are relatively unusual. The unusual bauxite can be investigated for potential impacts on refinery processes well in advance of mining.

The bauxite analysis technique is described conceptually. Challenges and difficulties are discussed. Results for an example application - the measurement of available alumina in Darling Range bauxite - are presented.

KEYWORDS:

bauxite, minerals, analysis, mining, FTIR, multivariate

Estimating bauxite quality USING FTIR Spectroscopy and multivariate calibration

Sharon L Eyer

INTRODUCTION

Bauxite deposits in the Darling Range, Western Australia (WA) vary greatly in chemical and mineralogical properties. On average, they are low in available alumina (34 %) but high in silica (21% SiO2) and iron (21% as Fe2O3). The Bayer process at the Kwinana, Wagerup and Pinjarra refineries in WA has been adapted to suit the low-grade bauxite. Nonetheless, excursions in bauxite quality can have serious impacts on process methods, equipment and raw material usage.

As knowledge of relationships between bauxite properties and refinery processes improves, supporting laboratories are asked to provide an increasing range of information about bauxite samples. A typical list would include some or all of the following.

  • available alumina, extractable alumina, total alumina
  • reactive silica, total silica
  • total iron, magnetic susceptibility
  • extractable organic carbon (EOC), extractable carbonate, extractable oxalate

The majority of standard analytical methods for these properties are technically excellent. However the resource requirements to provide the analyses are often considerable and the methods themselves can be time-consuming. Consequently, all-encompassing methods that produce many bauxite quality measurements are invaluable.

Exploration bauxite samples are usually analysed many years before the related region is likely to be mined. Laboratories often have to re-analyse samples for new properties that suddenly and unexpectedly become important to refinery operators. It is helpful to have a technique that can quickly provide such data plus information about historical levels in bauxite that has already been through the refinery process.

The RAMBO (Robotic Assessment of Mineable Bauxite Ore) system, developed and successfully installed at the Kwinana alumina refinery in Western Australia, meets all of these requirements. The benefits of the RAMBO system include the following.

  • High sample throughput – several thousand samples per day for 6-20 properties (limited only by computing power).
  • "Historical" analyses – new property information can often be derived for old samples without the need to re-analyse bauxite samples (computing task only).
  • Low manpower when fully automated – 1 to 2 analysts.
  • Minimal sampling handling – just provide bauxite in standard containers.

HISTORY OF RESEARCH AND DEVELOPMENT

At the heart of the RAMBO system is a series of models that relate bauxite properties to changes in data collected using Fourier transform infrared (FTIR) spectroscopy. Model development is a straightforward task with suitable software. However refining a model to meet analytical requirements can take weeks, months or years depending on experience and the underlying potential for the property of interest. A highly trained analyst, under the direction of a research chemist, currently undertakes the development of new models for the RAMBO system.

Table 1

The success of RAMBO system is the result of many years of research and development carried out by dedicated and highly skilled teams

Team

Period

Activities

Composition

Size

<1992

Idea conception, discussions with external researchers, purchase of FTIR instrumentation and software for multivariate calibration.

Experimental Scientist

<0.3

1992

Develop theoretical and practical understanding of technology. Highlight unique challenges when applied to bauxite.

Research Scientist

<1

1992-94

Build models. Analyse samples in parallel with established analysis techniques. Determine impacts and benefits for mine planning.

Research Scientist, Statistician, Geologists, Analysts

>5

1994-96

Assess feasibility of automation. Develop solutions to novel challenges including robot interaction with FTIR instrument, data handling issues, quality control strategies and robot handling of small envelopes.

Research Scientist, Statistician, Computer Programmers, Robot Experts, Geologists, Engineers, Analysts

>20

1996-97

Build RAMBO system. Develop calibration procedures and models for production system.

Successful installation in mining laboratory.

Research Scientist, Statistician, Computer Programmers, Robot Experts, Geologists, Engineers, Analysts

>30

Automation is required to make the system truly useful. The spectroscopic analysis is a surface technique and therefore sample preparation must be consistent. Automation is the best way to achieve consistency and speed. The RAMBO system not only handles and prepares bauxite samples, but also deals with the transfer of data to mining systems, monitors quality control aspects and provides long-term storage for FTIR spectral data. The seamless integration of over a dozen different computing systems required a highly skilled team drawing experience from many different fields.

Once models are developed and the infrastructure is in place, the number of properties that can be analysed simultaneously is limited only by computing power. The RAMBO system currently provides estimates for six different properties; this is expected to increase to 20 properties within the next few years.

INFRARED SPECTRA OF BAUXITE

At the outset, it is not necessary to fully understand or interpret the FTIR spectral data because understanding is often gained during method development. In some case, investigators can acquire potentially useful information about bauxite composition and the interactions between constituents. Experience and understanding can, however, improve the potential to produce successful models.

Infrared (IR) spectra contain information about the structure and composition of analysed samples. FTIR spectra are obtained by irradiating samples with a continuum light source then measuring the frequency and intensity of radiation absorbed by the molecular bonds within the samples. Many functional groups such as –CO, –COOH, –CH3, and –OH produce absorption peaks at characteristic frequencies. The intensity of absorption at such frequencies is usually proportional to the concentration of the relevant molecular group in the sample. Detailed investigation of FTIR spectra can lead to sample identification and enable quantitative analysis. Results can often be confirmed by comparing unknown spectra with spectra from certified samples.

IR spectroscopy is commonly associated with the analysis of organic materials but it is increasingly used to complement x-ray diffraction (XRD) for mineral characterisation. Characteristic absorption frequencies for some common bauxite minerals are shown in Table 2

Table 2

Some characteristic IR absorption frequencies for selected minerals. (Gadsen, 1975)

Gibbsite

Boehmite

Hematite

Goethite

Kaolinite

3685a

3617 s,spb

3520 vs

3445-28 vs

1030-18 vs

800 s

749-40 s

670-60 s,sh

585 m,sh

562-40 vs

515 vs

3290-45 s,sp

3095-40 s,sp

1160-35 b,sh

1085-65 s

760-30 vs

630-05 vs

525-00 vs

 

 

1175 sh

1160 sh

560-50 vs,b

532 b,sh

 

 

3095-2985 mb

1660

1105 m,b

1040 b,sh

912-882 vs,d?

812-793 vs,d?

672 m,sh

599-578 s,b

3696 s,sp

3670-56 m,sh

3645 w

3630-24 s,sp

1117-05 s,sh

1035-30 s

1019-05 s

940-35 m,sh

918-09 s

700-686 m

542-35 s

  1. Frequency units are wavenumbers (cm-1) and are listed in descending order.

(b) v – very; s - strong; m – medium; sp – sharp; b – broad; d – doublet; sh – shoulder

Figure 1

Approximate locations of the main characteristic absorption frequencies for (a) gibbsite (b) boehmite (c) hematite (d) goethite and (e) kaolinite highlighted on an FTIR spectrum for a Darling Range bauxite sample.

MODELS FOR ESTIMATING BAUXITE QUALITY

FTIR spectra are normally viewed as plots of absorbance against frequency. Bauxite spectra are challenging to investigate because the spectral features are broad and indistinct (Figure 1). Nevertheless, an experienced investigator can still extract useful qualitative and quantitative results because most of the necessary information is contained within the spectral data - though it is well concealed. Sophisticated statistical tools such as univariate and multivariate calibration are used to tease out the required information.

4.1 Univariate Techniques

It is theoretically possible to develop quantitative models for some bauxite properties without resorting to multivariate calibration. Absorbances at individual infrared frequencies (3455 cm-1 and 3530 cm-1) have been used to determine the gibbsite composition in selected Hungarian bauxite (Jonas and Solymar, 1970).

Similarly, available alumina in Darling Range bauxite can be estimated from a model based on absorption at a single FTIR frequency. Firstly, a set of samples is analysed for available alumina by the best available analytical technique. Available alumina in bauxite is normally estimated from the dissolution of the aluminium species during simulated bauxite digestion. The amount and type of aluminium species reporting as available alumina depends on the mineralogy of the bauxite, the refinery digestion conditions and the exact details of the laboratory method. For bauxite from the Darling Range, the digestion method usually measures soluble gibbsite, however boehmite may also dissolve and be included. (Roach, 1985).

Meanwhile, FTIR spectra are collected for the same set of samples. A relationship is then established between concentration and IR absorption at a single frequency. During method development, potentially suitable frequencies are determined from spectral libraries, if available, and by a systematic examination of all available frequencies. It is a simple task with a modern computer to calculate correlation coefficients or similar statistics for every available FTIR frequency. Frequencies that correlate strongly with concentration are modelled individually until the best calibration model is obtained.

Finally, the best calibration model is inverted and can be used for prediction. An example of the predictive ability of a univariate model is shown in Figure 2 (a). The frequency chosen is strongly correlated with available alumina (r = 0.86). There is a reasonable relationship between the data predicted using the FTIR-based model and the results collected using the best-known laboratory digestion technique. Significant refinement is required, however, before the model would be suitable for a practical application.

Unfortunately, there is limited scope to improve the univariate model. The main reason is that univariate models are often confounded by a myriad of other bauxite properties. For example, the frequency used in Figure 2 (a) has a strong negative correlation (= -0.69) with the concentration of reactive silica in bauxite, and statistically significant correlations with a number of other properties. Selection of a frequency free from interference is complicated, often impossible, because of the highly variable composition of bauxite and the complex interaction between constituents.

Therefore univariate prediction models developed for bauxite samples are extremely vulnerable to spectral and sample-related interferences. The risk of producing poor quality, unreliable, prediction data is very high. This in turn limits the potential for widespread application to bauxite mining or other refinery processes.

Figure 2

Prediction of available alumina from FTIR models calculated using (a) univariate calibration and (b) multivariate calibration. The Y=X line is shown on each plot. The X-axis data was obtained using a standard technique based on bauxite digestion.

4.2 Advantages of Multivariate Techniques

The main advantage of the multivariate technique for bauxite samples is shown clearly in Figure 2 (b). Here, a simple, unrefined multivariate model is used to predict the available alumina concentration for the same set of bauxite samples displayed in Figure 2 (a). The results are considerably more accurate than the univariate predictions and the model can be applied to practical applications. Similar improvements, not shown, are observed for all other bauxite properties analysed using the RAMBO system.

Multivariate models are developed in a similar way to univariate models except that more sophisticated statistical tools are used to build a relationship between FTIR data and bauxite properties. The primary reason for the accuracy improvement lies in the fact that the multivariate model incorporates every frequency in the supplied FTIR data. Frequencies strongly correlated with the property of interest are given a higher weighting than frequencies that are poorly or inconsistently correlated. Confounded bauxite properties are resolved or compensated by negative weightings on appropriate data points. Effectively, the noise in the relationship between the property and FTIR data is reduced markedly and the predictive ability is enhanced compared to the univariate model.

Understanding a second advantage of multivariate models – estimating the reliability of a predicted result – requires some knowledge of how multivariate models are calibrated. FTIR spectra are collected from several hundred bauxite samples selected as "standards". Statistical tools, usually principle component analysis (PCA) or partial least squares (PLS) routines are used to summarise the spectral data. Conceptually the techniques determine a set of lowest common denominators for the calibration samples, which are often visualised as abstract spectra.

 

Factor 1: F1(u ) = { F1(u 1) , F1(u 2) , F1(u 3) ,…, F1(u n) } (1a)

Factor 2: F2(u ) = { F2(u 1) , F2(u 2) , F2(u 3) ,…, F2(u n) } (1b)

Factor 3: F3(u ) = { F3(u 1) , F3(u 2) , F3(u 3) ,…, F3(u n) } (1c)

Calculated Spectrum A (u ) = l1A F1(u ) + l2A F2(u ) + l3A F3(u ) +…+ lmA Fm(u ) (2a)

Calculated Spectrum B (u ) = l1B F1(u ) + l2B F2(u ) + l3B F3(u ) +…+ lmB Fm(u ) (2b)

Calculated Spectrum C (u ) = l1C F1(u ) + l2C F2(u ) + l3C F3(u ) +…+ lmC Fm(u ) (2c)

  Where u n is the nth wavenumber frequency and lFa is the spectrum specific score for factor F.

The outcome is a series of factors (1) that can be used to mathematically reconstruct any spectra from any individual sample (2). The important aspect in this scenario is that there are now two spectra produced for each sample – one calculated spectrum and the original measured FTIR spectrum. The quality and reliability of a prediction result for any sample can be estimated by comparing these two spectra. Samples that are similar to those used to generate the model usually produce spectra that are easily reconstructed from the factors. That is, the difference between the measured and calculated spectra is small. Conversely, samples with poorly reconstructed spectra often have characteristics that are atypical of the samples in the calibration set. Subsequently, predicted results for atypical samples are treated with suspicion.

The type of samples that are highlighted as unusual depends upon the range of samples in the calibration set and the root cause of the spectral variation. The quality estimate is not exclusive and it is possible for some poorly predicted samples to have well-calculated spectra and for some poorly re-constructed spectra to be accurately predicted.

Nonetheless extreme cases are always detected and the ability of the multivariate techniques to automatically provide such information is more than a significant convenience. If the composition or nature of a bauxite sample is outside the normal range for one or more properties then there is a high chance that it will be highlighted as unusual. This is true even if the property causing the uncertainty is not specifically measured. Potentially unusual samples highlighted in this manner can be investigated for potential impacts on refinery processes.

4.3 Challenges with Multivariate Techniques

The most challenging aspect of developing multivariate models for bauxite is establishing a suitable suite of samples for calibration. The range of calibration samples in terms of nature, composition and source must cover the range expected from future unknown samples. There is always a compromise between robust models and accurate ones. Models for exploration bauxite are calibrated using a wide range of samples so that they are very robust, although there is some loss of accuracy. A narrower calibration range can produce more accurate predictions for some samples but will be far more susceptible to errors for even slightly unusual samples. Thus the average prediction quality for a region of exploration bauxite, which by definition is highly variable in nature and composition, is usually lower with a narrowly targeted calibration set.

Due to the manner in which multivariate FTIR models are developed, they are fully reliant on data from reference analytical methods. Therefore, the models can never be more accurate or more precise the reference techniques when compared on a one-to-one basis. Model precision can often be improved, however, by averaging the prediction results from a larger number of samples than could be reasonably analysed using the reference analytical technique.

DISCUSSION

The multivariate models developed for the RAMBO system are remarkably accurate and extremely robust. Prediction data can be used with a high degree of confidence. This is enhanced by the ability to automatically highlight potentially unusual samples, which can then be set aside and investigated in detail for potential impacts on refinery processes. This is an additional benefit, not often available when using standard, laboratory, analytical techniques.

The FTIR spectral data used to build the multivariate models incorporates a great deal of chemical and physical information about the samples. Multivariate models developed from a set of bauxite samples take into account normal variations for all properties – including those that are not measured or that may not be known. Consequently the FTIR spectral data, if stored appropriately, can be retrieved and re-analysed any number of times for new properties as required.

The hidden implication is the ability to rate the "normality" of a bauxite sample compared to previously analysed samples. Geologists are able to use both direct and indirect property data from the RAMBO system to provide consistent bauxite supply to the alumina refineries. Since excursions in alumina content and bauxite impurity levels can cause problems with refinery process control, the potential for improvements in refinery efficiency and product quality is clear (Roach, 1975).

Conclusions

The RAMBO system can replace many reference techniques that meet refinery requirements in terms of data quality but require large resources and have limited throughput. The system combines multivariate models with automated sample handing to provide a fast, accurate and robust technique for simultaneously estimating a large number of bauxite quality parameters. The range of applicable properties includes those that are well defined and understood and (potentially) those that are yet to be identified.

The capacity to obtain a greater understanding of the geology and mineralogy of the bauxite resource is yet to be fully explored.

ACKNOWLEDGMENTS

The author is grateful to many people who have contributed to the work reported in this paper. Particular thanks go to Dr. S. Grocott, Dr. D Mason, Dr. G. Riley, G. Hancock and to all of the analysts in the Kwinana Mining Department.

 

notation

u = frequency of IR radiation = 1/l

Wavenumber (cm-1) = frequency unit commonly used for mid-IR spectroscopy (cm-1)

l = wavelength (cm)

Xwn = multivariate coefficient n for wavenumber u

lnF = multivariate loading n for factor F.

references

Gadsen, J.A., (1975), Infrared Spectra of Minerals and Related Inorganic Compounds, Butterworth & Co., ISBN 0-408-70665-1.

Jonas, K. and Solymar, K. (1970). Determination of the mineral composition of bauxites by infra-red spectrophotometry. Acta Chimica Academiae Scientiarum Hungaricae, 66(1), pp1-11.

Roach, G.I.D and Prosser, A.P. (1975).Australian I.M.M. Conference.

Roach, G. I. D. The Influence of mineralogy on the dissolution kinetics of gibbsite. Light Metals 1985 New York: American Institute of Mining, Metallurgical, and Petroleum Engineers, pp183-196.