In every instance of process or laboratory measurement it pays to understand the source of errors and to minimize those errors where possible. From the sampling system to the instrument or analyzer settings, to the method development used for interpreting the signal, to the maintenance of the calibration over time, all have something to contribute to the error profile. In the case of optical spectroscopy, care needs to be exercised in the initial setup and data processing method development, but once set, these remain constant for the life of the system. Minimizing error then falls to the routine calibration maintenance, which requires constant monitoring of the process results.

In Part 1, one suggestion was to improve the precision of the reference method by running duplicate samples and averaging or adopting the median value. However, we can take advantage of the many samples that are collected during the calibration process, and averaging may not be necessary. Consider a typical calibration as shown in the figure. The distribution of points around the regression line is controlled by errors both in the spectrometer/sampling system and in the reference method. In a gaussian distribution, the best answer is at the apex. If we rotate the regression line to be oriented vertically, we essentially are mapping the points to a gaussian. This means that the best estimate is anywhere on the regression line.

Here is the link to the full presentation on Minimizing Error in Calibrating Spectrometers from ATC 2025 Conference. Also available in Youtube video from 00:00 to 22:00.

Choosing the best method will optimize how future spectra will be processed. The calibration procedure for spectroscopic models begins with selecting appropriate preprocessing methods. Different analyzer technologies (e.g., NIR vs. Raman), spectrometer manufacturers, and chemometric algorithms can influence the optimal choices. To evaluate preprocessing methods, Root Mean Squared Error of Cross Validation (RMSECV) is commonly used, offering a reliable measure of combined bias and precision error.

In one example, various combinations of preprocessing techniques (e.g., derivatives, scatter correction, normalization) were tested. The results showed that some combinations significantly outperformed others, with the best approaches selected based on minimizing RMSECV with the fewest model factors.

Partial Least Squares (PLS) regression is the standard algorithm due to its widespread integration and effectiveness in handling errors from both spectroscopy and lab sources. However, PLS assumes linearity, which doesn’t always apply. In such cases, alternative methods like Locally Weighted PLS (LWR-PLS) can provide better performance by addressing non-linearity through localized modeling.

Finally, method development is a one-time effort. Once a spectral data processing method is chosen, it generally remains fixed throughout the instrument’s use.

Here is the link to the full presentation on Minimizing Error in Calibrating Spectrometers from ATC 2025 Conference. Also available in Youtube video from 00:00 to 22:00.

The terms “accuracy” and “precision” are often confusing. Accuracy refers to how close a result is to a known value, while precision indicates the consistency of repeated results. Ideal measurements are both accurate and precise. However, it’s possible to have high precision with poor accuracy, or low precision that still averages out to an accurate result over many trials.

Application to Octane Measurement:
In octane rating analysis, precision is a challenge, especially with the reference octane engine, which is not always precise despite being used to define accuracy. Essentially, the octane engine represents the “Not precise, maybe accurate” portion of picture. Repeated measurements of the same gasoline sample show a normal distribution of values with a standard deviation of ~0.25 octane units, meaning 95% of results fall within ±0.5 units.

The calibration process, then, is charged with mapping a very precise spectrum to a far less precise reference value and calibration must account for this. One can run multiple engine tests on the same sample and use the average value to reduce error—error decreases with the square root of the number of runs. In dealing with a system where the reference is imprecise, there is more that we can do.

Here is the link to the full presentation on Minimizing Error in Calibrating Spectrometers from ATC 2025 Conference. Also available in Youtube video from 00:00 to 22:00.