Stan: Why isn't NIR and Raman your first choice?
Jim: These analyzers are much more difficult to understand and troubleshoot. I can kind of explain the function of an FTIR, IR and UV analyzer to technicians in terms they can easily relate to, such as a prism separating the colors of light, a microwave cooker, etc., but when I start to talk about statistical models and billions of calculations in a NIR or Raman analyzer, eyes glaze over. The technician is almost entirely dependent upon the supplier for technical support and the analyzer specialist, as well as model support if the PLS models were developed by the supplier. The development of these models requires special expertise despite software claims to the contrary that all you need to do is feed data to get wonderful results. A graduate degree in statistical methods is often a prerequisite.
Greg: Besides the maintenance issue, what about performance?
Jim: As with any statistical model, the prediction is only as good as the comprehensiveness of the samples used to develop the model. Often production units are not allowed to be deliberately run to cover entire spectra of possibilities, particularly abnormal conditions. The plant's data is also not truly random and has patterns and internal correlations, often from closed-loop control. Lab samples can be prepared, but differences in temperature and unknown components can invalidate models. Cross-correlated or coincident inputs can lead to erroneous relationships. The inputs (components) may not be independent variables. Recently the use of samples with a spike of the pure component of interest has been recommended to break offending autocorrelations. A PLS model is indicative of a correlation between the components and spectra, but not a guarantee of cause and effect.
Unlike the use of (PLS) models for production unit analysis, there are no first principles that can be readily be used to evaluate whether the identified correlations are deterministic or accidental. It may take years to develop enough data that covers all of the possibilities and to gain confidence in the results. For example, a NIR used on a tank farm took several years to be valid due to variation in seasonal temperatures, but could still be fooled by an exceptionally cold winter. I have been burned by an NIR that said the process was running great when in reality it was running amok. Generally, the use of NIR or Raman analyzers developed for your proprietary concentrations should only be used for advisory or restricted trim control until there are more than two years of valid results. This is not to say NIR is a last resort, but rather a choice when more selective techniques are not available. The cost of a GC is about the same as NIR. A sample system is usually required for a GC and may be required for NIR to prevent coated probes or noisy spectra, but the sample system design is generally less extensive. However, the maintenance cost of a NIR can easily exceed any savings in the sample system. Mass spectrometers can be more costly. The motivation as to analyzer selection is usually not as much cost-related, but is what works. We used NIR to help monitor and control polymerization condensation reactions and some blending applications.
Stan: Where has NIR been most successful across the process industry?
Jim: NIR is the standard for octane analyzers because of the maturity, generic nature and definitiveness of the application. If there is a big enough market, the analyzer manufacturers will develop robust models, updating them as necessary.
Greg: For the prediction of corn fermentability and, hence, essentially yield, a near-infrared transmittance (NIR-T) analyzer has been developed by Monsanto that enables a reduction in the carbon footprint and corn use that is more than 50% of ethanol production cost. Since ethanol is obviously not present in the corn, and the first principle relationships between starch and nutrient content and the conversion to sugars and ethanol are not sufficiently defined, a statistical model is an obvious choice. Monsanto updates the model yearly and freely provides them to ethanol producers to make the industry more efficient. I developed a simple production rate controller that immediately reduces corn feed rate to take advantage of an increase in predicted fermentability using the enhanced PID developed for wireless (See Control, May 2009, p. 54, Is Wireless Process Control Ready for Prime Time?.)
A running average of batch time to ethanol end point is used to automatically correct the predicted ethanol yield online.
Stan: In future months we cover other analyzers, such as viscometers, microwave and nuclear magnetic resonance, get a perspective of emerging opportunities in batch and continuous processes, and look at how to revitalize end-user analyzer expertise.
Greg: We conclude with a poignant Top Ten List.
Top Ten Reasons Not to Buy an Analyzer
(10) Don't want to upset your friends in the lab by second guessing them
(9) Some things are better left unknown
(8) Analyzer specialist looks awfully old
(7) Can't toss analyzer in the back of a pickup truck
(6) Like playing "Who Dun It?"
(5) Temperature trends are a lot smoother than analyzer trends
(4) Don't want some fancy new advanced control system
(3) Don't need real-time optimization; we are running the best we can
(2) Cool new software can predict everything by just feeding it all your data and pressing a button
(1) Consulting firm says if you stop buying analyzers, you can afford bigger executive bonuses.