Analysing Raman spectra of crystalline cellulose degradation by fungi using artificial neural networks

Sammanfattning: This thesis investigates the use of artificial neural networks for classifying Raman spectra of partially degraded cellulose samples by fungal species. A multilayer perceptron configuration of 4 hidden layers and 128 hidden nodes was able to classify a set of 60 samples with an overall prediction accuracy of 0.55. Results show that data resolution is an important factor when optimizing classifier performance, and that a resolution of 1.0 cm^(-1) gave the highest performance. We found that choosing suitable parameters for the asymmetric least squares smoothing (ALSS) correction is of relevance when attempting to optimize classifier performance, and that an ALSS smoothness value of lambda = 10^5 gave the highest performance. Results also indicate that some fungal species and control treatments have stronger signatures in certain spectral regions. Gloeophyllum sp., Coprinellus angulatus and NaOH treatments had the most accurate probability distribution and may therefore be considered to cause the most unique cellulose modification. This thesis shows promising results for artificial neural networks to be utilized for classifying Raman spectra of partially degraded cellulose samples.