The Application Of Machine Learning In Analyzing Atomic Spectroscopy And Molecular Spectroscopy

Spectroscopy technologies have a lot advantages,such as fast-detecting,pollution-free,and no need for complex sample preparation.However,there are some problems in the spectral analysis,such as the spectrum is susceptible to the matrix effect,temperature drift,light source energy jitter and other experimental environments,so the spectral stability is poor.By applying machine learning to the spectral analysis process,many factors that are not conducive to spectral analysis can be overcome,including the ability to extract more information from the spectral data,reducing the effects of interference signals,and detecting outlier data.This paper introduced five machine learning algorithms,namely Principal Component Analysis(PCA),Partial Least-Square(PLS),Random Forests(RF),Deep Neural Networks(DNN),and Convolutional Neural Networks(CNN).The above algorithms applied in atomic spectroscopy-Laser Induced Breakdown Spectroscopy(LIBS)and molecular spectroscopy-microscopic infrared spectroscopy for spectral feature extraction,noise reduction,model construction,etc,were discussed,and further evaluated the effect of machine learning algorithms in the process of spectral quantitative analysis and qualitative analysis.At first,the study took the analysis of Chemical Oxygen Demand(COD)as an example,and applied machine learning to establish a LIBS-based rapid prediction model of COD.COD is an important indicator of water pollution,and the result of a combination of multielements.Therefore,PLS,a multivariate regression algorithm,was used to establish a quantitative model,but the modeling ability of different rivers was poor.The reason was that the elements in the samples of the two rivers are significantly different,which leads to the poor transferability of the linear PLSR model.Aiming at this problem,this paper studies the quantitative model of COD based on random forests(RF).By optimizing the parameters of the random forest,the model based on RFR is better than PLSR.This case showed that machine learning applied in atomic spectroscopy successfully.Further,this paper also studied the application of machine learning applied in molecular spectroscopy.Taking the infrared microspectroscopy in molecular spectroscopy as an example,the origins and species of rice were distinguished as a case study by machine learning.The experiment collected 14 kinds of rice's infrared microspectroscopy,using the regular machine learning-RF to distinguish the origins and types of these samples,but the accuracy in test set of the model was only 57.1%.In order to accurately distinguish the origins and varieties of rice,this paper also used two deep learning algorithms(DNN and CNN),and established models for distinguishing origins and types of rice,respectively.The deep learning algorithms gave the best accuracy of test set and the accuracy of both deep learning models is over 90%.The results showed that infrared microspectroscopy combined with deep learning is a faster and more effective method than regular machine learning in tracing origin and identifying of rice,which also showed that machine learning applied in molecular spectroscopy successfully.The above problems all involved a large amount of data processing.They have been properly solved by applying machine learning by applying machine learning in processing atomic spectroscopy and molecular spectroscopy,which shows that machine learning is a very effective tool in spectral analysis,also shows that machine learning algorithms have advantages in both quantitative and qualitative analysis.Furthermore,machine learning can be extended to the application of spectral data analysis and modeling of other objects.