Machine learning is able to reliably extract an incredible amount of data from both Raman and LIBS spectra. One of the most interesting recent examples was published recently in Applied Spectroscopy , in which Julio Banquet-Terán and co-authors showed that a combination of genetic algorithms combined with a backpropogating neural network was able to make excellent predictions from Raman spectra of polyproplene taken at 785 nm. Five Raman spectral accumulations of 10 seconds each were used to compare properties measured with the ASTM standard methods for Youngs traction modulus, tensile strength at yield, elongation at yield on traction, and flectural modulus at 1% secant.
Numerous types of polypropylene were used, including homopolymer PP, impact ethylene-propylene copolymer, and ethylene-propylene copolymer. Among all of these samples, validation data (taken after calibration with separate data) showed only a relative standard error of between 0.42 and 0.88% for the various predictions. The interesting thing is that these Raman measurements are far faster than the traditional ASTM methods, and they are nondestructive as well. If you’d like to look into implementation of these methods in your plastics process, please contact us.
This is yet another example – similar to coal heating value and volatile content with LIBS – of material properties being predicted with broadband spectra using multivariate methods. If you’re interested in implementing machine learning with your spectral sensor, now is the time. The tools and the implementation understanding are progressing rapidly!
To read more of this article, see Julio Banquet-Terán, Boris Johnson-Restrepo, Alveiro Hernández-Morelo, Jorge Ropero, Miriam Fontalvo-Gomez, and Rodolfo J Romañach, “Linear and Nonlinear Calibration Methods for Predicting Mechanical Properties of Polypropylene Pellets Using Raman Spectroscopy,” Appl. Spectrosc. 70, 1118-1127 (2016)