Exploring ToF-SIMS Data Analysis with Spectral Binning Using Multivariate Analysis and Machine Learning

Submission note: A thesis submitted in total fulfillment of the requirements of the Degree of Doctor of Philosophy (PhD) to the Centre for Materials and Surface Science (CMSS), Department of Chemistry and Physics, School of Molecular Sciences, La Trobe University, Victoria, Australia. In Conjunction with Commonwealth Scientific and Industrial Research Organization (CSIRO) Clayton, Victoria, Australia.

The advancement of surface analytical instruments and data analytical capabilities is extremely important for understanding increasingly complex samples and materials. Time-of-flight Secondary Ion Mass Spectrometry (ToF-SIMS) is a key technique capable of deciphering these complexities and gaining insight into the elemental, isotopic and molecular composition and structure present at the sample surface. The increased complexity and scale of data available necessitates the use of advanced computational methods evaluate the acquired data. In this thesis, we investigate the use of spectral binning on datasets of varying complexities from discrete polymers to biomaterials. All spectral data are assigned to simplified data bins that generate a characteristic ‘fingerprint’ of each spectrum acquired. Spectral data are also presented in the form of manually and automatically selected peak lists to allow a broad range of analytical opportunities and for comparison with mass segmented datasets. Multivariate analysis (MVA) and machine learning, in the form of self organizing maps (SOMs), are deployed to interrogate these datasets. Principal Component Analysis (PCA) and other MVA tools have drawbacks with large datasets and are skewed if any anomalies are present in the acquired data. This thesis demonstrates that machine learning overcomes these anomalies, by emphasizing the similarities between samples and minimizing the differences caused by these anomalies. The customizable machine learning algorithm constructs a computational model on the input spectra and highlights the distinguishing features that clearly identify each analysed sample. Furthermore, optimization and extension of spectral binning and understanding the information content present in ToF-SIMS spectral data are explored, demonstrating the applicability, versatility and robustness of pairing spectral binning and SOMs. This thesis exemplifies the usefulness of spectral binning and machine learning for ToF-SIMS analysis. It is applicable beyond ToF-SIMS analysis with the sole requirement of forming a data matrix that encompasses all acquired data, irrespective of information complexity.