Machine Learning Microplastic Characterisation Surpasses Human Performance and Uncovers Labelling Errors in Public FTIR Data

Abstract Microplastics are ubiquitous and appear to be harmful, however, the full extent to which these inflict harm has not been fully elucidated. Analysing environmental sample data is challenging, as the complexity in data makes both automated and manual analysis either unreliable or time-consuming. To address challenges in data analysis, we explored the use of a dense feed-forward neural network on a diverse dataset of Fourier transform infrared spectroscopic data. The model makes predictions that are probabilistic multi-category vectors, emulating onehot encodings for each of the classes, thereby indicating the probability of a given sample belonging to a known class. Our results indicate that this model outperforms human classification performance for environmental microplastic FTIR data. This is assumed as the model broadly reproduces human decisions, while also revealing systematic errors in human interpretations. These uncovered errors indicate that for a model making informed and reliable decisions, there exists an artificial limit to the realistically achievable performance in metrics, where metrics measure agreement between human and model predictions. This work indicates an enormous potential for a small and highly efficient dense feed-forward neural networks in making reliable and fast analysis of large volumes of complex FTIR data accessible.