Vendor column: Creation of a quantitative data set

Skip to Navigation


  • Published: Oct 6, 2016
  • Author: Daniel Shiley
  • Channels: Infrared Spectroscopy / Chemometrics & Informatics
thumbnail image: Vendor column: Creation of a quantitative data set

Daniel A. Shiley – Senior Application Chemist, SummitCAL Solutions Team, ASD Inc.

This is the third of a four-part series on chemometric applications for materials analysis.

In my previous articles, I have discussed the use of near-infrared (NIR) for production of qualitative mineralogy of mining-related samples. Qualitative models are best suited to the exploration or early production phases, but once the mine begins operation a quantitative mineral model will provide the most useful information to aid in process optimization. In mining applications, the term minerals describe specific mineral species such as kaolinite and not mineral elements such as iron or phosphorus. Although mining companies have been using near infrared (NIR) for years, few have developed true quantitative prediction models.

Creation of a quantitative data set for mining—Measuring minerals 101

Last month we learned about creation of alteration maps or mineral maps with NIR. These predictions and estimations yield important information to help identify areas that may contain the target metal. Qualitative estimations only identify what they can see, in other words, if spectral signatures were detected for three minerals, these three minerals would comprise 100% of the total mass. We know that using a qualitative mineral model results in values that do not relate to laboratory quantitative measurements of the minerals because many minerals such as pyrite and quartz do not have spectral features. For example, if the actual weight percentage of the sample included 45 percent quartz, 20 percent kaolinite, 20 percent muscovite and 15 percent pyrite, a measurement with NIR using a qualitative model would yield a result of 50 percent kaolinite and 50 percent muscovite. Although the percentage of mass is not correct, the presence of two important alteration minerals is correct and thus could be used to vector an area that may contain metal. This approach works well for exploration and early production planning, but begins to have some serious drawbacks once the mine begins operation. Engineers that are operating the ore processing plants need to know exact quantities of certain minerals that could cause problems with extraction of the metal or that increase the costs of operations.

Data set development for creation of the multivariate model – The first step towards quantitative analysis

In order to begin a quantitative project, a data set of spectra and reference analyses should be created. This process is really not much different for inorganic mineral samples than it is for organic materials. The task begins with a series of questions that should be answered.

  1. What sample type do we wish to measure?
  2. What constituents do we want to measure?
  3. Is reference data available for those constituents and sample type or will it need to be developed?
  4. How can we obtain representative samples for the deposit?
  5. What is the budget for this measurement project?

Sample type

There are many options for measurement of mineral (rock) samples. These include core, coarse chips and assay pulp. Solid core sample (which resembles a cylinder of solid rock) is not uniformly homogenous and so trying to use this sample type for quantitative measurement will result in higher error when the sample is measured by reference assay. Coarse chip samples also can be rather inhomogenous, so if using chips, it is necessary to measure a very large sample size. Pulverized samples or assay pulps are more homogenous, but if the particle size of these pulps becomes too small, this will result in reduction of the spectral features. Figure 1 contains a plot showing one sample prepared to three different particle sizes. FC is around 1 mm, 150 mesh is 100 micron and 400 mesh has 38 micron maximum particle size. In general, the finer the particle size the greater the overall reflectance value and smaller the wavelength features. Very fine pulverization can even eliminate certain wavelength features that were the result of the crystal structure of the mineral.

Figure 1. Spectral characteristics at three particle sizes.

Figure 1. Spectral characteristics at three particle sizes.

Additional considerations of sample particle size

Since varying particle sizes create spectral changes, an important consideration for a mineral analysis project should include some thought about particle size. We know that NIR can be rather sensitive to differences in particle size. Typically when working with organic materials, we simply grind the samples using a grinder that has screen openings of a certain size to constrain the upper bounds of the particle size. However, such a grinder does not exist for grinding rocks, so after crushing, the sample is sieved using a screen to remove the large, uncrushed particles. It is important to crush the portion of the sample that remains on the sieve and return the crushed portion to the sample so that the sample representivity is maintained. Not all minerals in rocks are equally hard, so particle size distribution of the sample may change depending upon which minerals are present. The crushing technique should be selected on the basis of which provides the most uniformity of particle size from sample to sample. Although when we create models we can utilize scatter correction to mitigate the effect of particle size, it is optimal to have uniform particle size distribution for the calibration sample set.


As previously discussed, not all minerals produce a spectral signature in the NIR region. So we typically recommend a review of reference library spectra to determine whether the mineral of interest has unique spectral features. The concentration of the mineral and the range of concentration of the mineral in the data set should also be evaluated. The constituent range should be as robust as practical.

Reference data quality – Trust everyone, but always cut the deck!

As with any calibration development project, the quality of the reference data is very important. If the reference data set is not of good quality, the NIR model will suffer, or may not even be possible. When beginning a project it is always a good practice to submit some samples to the laboratory in blind replicate so an accurate estimate of the reference error can be obtained. The samples submitted in blind replicate can also be very useful throughout the project, especially if the project spans a long period of time. A portion of these same samples can be included each time new samples are sent to the lab. This can help to provide an indication of the reference method stability. Unfortunately, laboratory methods often are changed and little attention is paid to the fact that each small change in the method also can cause a resulting change in the reference data produced by the method. Comparison of the same reference samples will help to identify when a change has occurred in the reference method. Of course, it is desirable to use same laboratory throughout the project because the differences between labs will likely increase the overall calibration error.

Obtaining representative samples

When beginning a quantitative project, it is a good idea to first check the archive for samples that have already had reference analysis. If the particle size of the archived sample is consistent with the project goal, these would be suitable to begin. It makes sense to locate samples that may be suitable since mineralogy reference analyses are expensive and time-consuming to obtain. Geologists can also use their knowledge of the deposit to identify regions that may produce the most diverse sample set. Generally, it is not a good idea to collect samples off of a conveyor because these likely have been blended during the transportation or grinding process such that they will tend to be more average than if samples are obtained from discrete locations.

Mind your budget!

While it is easy to suggest that more samples are better for creation of the calibration database, we always need to keep in mind that most projects do have a budget. Reference mineralogy can cost from $150 per sample to over $500 per sample depending on the type of reference assay. This adds a budget concern that should be addressed to begin the project. We suggest that the starting database should contain 150 to 200 samples. More samples could be obtained initially, but frequently the problem is that many of the samples that are sent to the lab could be similar enough that they do not help the calibration. So, in practical terms we want to have enough samples to represent the largest portion of the sample diversity that would be expected, but it is unrealistic to try to find every minor type that may be present in the initial data set. We will discuss strategies for improvement of the database in future articles.


Qualitative and quantitative determination of mineralogy is possible with NIR. Qualitative determination is used to identify important pathfinder minerals, whereas quantitative determination of mineralogy is used to optimize ore processing. Careful creation of the sample and reference database is the first step preparing to develop quantitative mineralogy with NIR.


  1. Goetz, A.F.H., Curtiss, B., Shiley, D.A., 2009. Rapid gangue mineral concentration measurement over conveyors by NIR reflectance spectroscopy. Minerals Engineering 22 (2009), 490-499.


Article by Daniel Shiley

The views represented in this article are solely those of the author and do not necessarily represent those of John Wiley and Sons, Ltd.

Social Links

Share This Links

Bookmark and Share


Suppliers Selection
Societies Selection

Banner Ad

Click here to see
all job opportunities

Copyright Information

Interested in separation science? Visit our sister site

Copyright © 2017 John Wiley & Sons, Inc. All Rights Reserved