Integration of machine learning techniques in chemometrics practices

Food safety is a key objective in all the development plans of the European Union. To ensure the quality and the sustainability of the agricultural production (both intensive and extensive) a well-designed analysis strategy is needed. Climate change, precision agriculture, green revolution and industry 4.0 are areas of study that need innovative practices and approaches that aren’t possible without precise and constant process monitoring. The need for product quality assessment during the whole supply chain is paramount and cost reduction is also another constant need. Non targeted Nuclear Magnetic Resonance (NMR) analysis is still a second-choice approach for food analysis and monitoring, one of the problems of this approach is the big amount of information returned. This kind of data needs a new and improved method of handling and analysis. Classical chemometrics practices are not well suited for this new field of study. In this thesis, we approached the problem of food fingerprinting and discrimination by the means of non-targeted NMR spectroscopy combined with modern machine learning algorithms and databases meant for the correct and easy access of data. The introduction of machine learning techniques alongside the clear benefits introduces a new layer of complexity regarding the need for trusted data sources for algorithm training and integrity, if this kind of approach proves is worth in the global market, we’ll need not only to create a good dataset, but we’ll need to be prepared to defend against also more clever attacks like adversarial machine learning attacks. Comparing the machine learning results with the classic chemometric approach we’ll highlight the strengths and the weakness of both approaches, and we’ll use them to prepare the framework needed to tackle the challenges of future agricultural productions.