Overcome the obstacle of NP analysis – a concept of chemical/microscopic methods combined with artificial intelligence

Numerous studies have shown the potential risk that nanoplastic (NP) represents for the living organisms in the different ecosystems. However, the amount and characteristics of NP present in the environment are still unknown in its full extent. Even if several methods have already managed to quantify or characterize environmental NP, none, to our best knowledge, could yet provide a single particle complete characterisation over the full nanoscale range combined with a high sample throughput.The present work tackles the challenge of NP full characterisation in soil by testing an innovative combination and alignment of µ Raman spectroscopy (RS), scanning electron microscopy coupled with energy dispersive x-ray spectroscopy (SEM/EDX), pyrolysis gas chromatography mass spectrometry (Py-GC/MS) and artificial intelligence (AI). The aim is to use the RS data to train an AI model that can automatically recognise NP in environmental samples using SEM/EDX data. The SEM data used to classify the particles include textural features extracted from the 2D images, elemental composition given by the EDX spectrum and the particle behaviour under the electron beam. Particles shape/size transformation when being exposed to high voltage has already been used for microplastic identification but still need to be tested for NP.First, NP down to 500 nm are identified using RS in samples of increasing complexity, starting with pure NP, mixed NP, spiked media and, finally, environmental samples. Secondly, the suitability of NP behaviour under electron beam to identify plastic material in complex matrices with SEM is tested on the identified NP. Then, the dataset acquired with RS and SEM/EDX on NP is divided into a training and testing set to build a convolutional neural network (CNN) allowing the differentiation between NP and non-NP particles present in a sample. Finally, textural features, elemental composition and behaviour under the beam data are acquired for all particles down to 50 nm in the different samples. The total mass of each polymer present in the sample is extrapolated and then cross-validated by performing a Py-GC/MS analysis on the same sample. Monte Carlo simulations are then used to model the error of the extrapolation based on data provided by the RS and SEM data. The aim of this model is then to allow the identification and characterisation of