We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Application of MATLAB and SAS Viya AI Models towards the Elucidation of Potential Microplastics in the Neuse River Basin
Summary
Researchers collected 18 water samples from the Neuse River basin and applied ATR-FTIR spectroscopy combined with principal component analysis, K-means clustering, MATLAB Classification Learner, and SAS Viya machine learning models to identify weathered microplastic particles whose spectra are obscured by environmental degradation. The multi-model approach improved identification accuracy by comparing against both a 7-polymer and a 9-polymer reference library.
Abstract Microplastics (MPLs) are ubiquitous particles derived from degradation from plastic refuse or directly from anthropogenic sources. These particles are present in aquatic environments with potential toxicology across the biosphere. In order to characterize their presence, the Neuse River basin was selected for sampling. However, like many MPLs, spectroscopic characterization can be sullied by weathering process which obscure the native spectra. Therefore, to enhance MPL identification, Principal Component Analysis (PCA), K-means Clustering (KMC), MATLAB, and SAS Viya’s machine learning (ML) modules were implemented on Neuse River basin samples. 18 unknown samples were recorded with attenuated total reflectance-Fourier transform infrared (ATR-FTIR) spectroscopy against the controls: high-density polyethylene (HDPE), low-density polyethylene (LDPE), polypropylene (PP), polystyrene (PS), polyvinyl chloride (PVC), polyethylene terephthalate (PET), and nylon-6 (PA6). Later on, these unknowns were tested against a new set of 900 commercial control samples partitioned by 9 classes: cellulose acetate (CA), HDPE, LDPE, PP, PS, PVC, PET, polymethyl(methacrylate) (PMMA), and PA6 using µ-FTIR (micro-FTIR) for single-particle discrepancy. Application of MATLAB’s Classification Learner and SAS Viya for Learners software helped aggregate multiple machine learning (ML) models to test for corroboration of predictions with datasets derived from Version 1 (V1), Version 2 (V2), and Version 3 (V3) feature extraction algorithms. Novel feature extractions, based on regressions of discrete data from spectral intensities, in V2, and unique probing of the relationships informing the regressions, in V3, showed moderate to moderately high predictor strength contributing to accuracy increase in V3 according to various predictor strength algorithms in MATLAB. In the test scenario of the strongest version, V3, 63.2% (+5.3% from V1) of the models performed very strongly (90% cutoff in accuracy) and 89.5% (+0% from V1) of the models performed moderately strongly (80% cutoff in accuracy). The models, coupled with the unsupervised PCA and KMC, indicated microparticle (MP) from Stockinghead-Creek in Duplin County, NC; SHR-1b(2), as LDPE. However, the change in corroboration across model types was not significant according to Kruskal-Wallis (H = .555, p = .7577). An ANOVA was performed to see if LDPE incidence was similar across all unknowns indicating a similar ratio of predictions, excluding NRCP-1 (F = 31.479, p < .0001). While it is unclear if this particle is truly LDPE, the results may suggest that LDPE could be of high presence in the river basin (specifically, PE) when taking into account the predictor set across all versions.