<b>Real-Time Classification of Mixed Plastics Using Machine Learning Algorithms Based on THz-TDS and Raman Spectroscopy Data</b>

<p dir="ltr">The effective recycling of used mixed plastics is critical to various industries worldwide that aim to promote sustainable practices. A fundamental component in this process is the sorting or classification of material in plastic streams prior to the fabrication of new products. Errors in this step could lead to chemically incompatible plastics in the same streams, leading to highly immiscible plastic mixes and a degradation in final product quality. Due to the increase in complexity of plastics with additives and fillers, current methods have proven to be ineffective for large-scale automated sorting. This thesis focuses on developing a novel sorting method using a combined Terahertz time-domain spectroscopy (THz-TDS) and Raman spectroscopy system that employs a machine-learning-based classification algorithm for the real-time classification of mixed plastics. Six types of mixed plastics were selected from a collection of industry-provided polymers that are relevant in various manufacturing contexts. A library of THz and Raman spectra was generated for the selected polymers through in-lab measurements. Dimensionality reduction was performed on the combined spectra using Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) to accentuate crucial spectral features and enable real-time classification through increased processing speed. A machine-learning (ML) algorithm was trained using this dataset for real-time polymer identification. This algorithm achieved over 99% predictive accuracy via class selection and 90.3% predictive confidence averaged across all polymers tested. It was found that the method of dimensionality reduction is crucial to the effectiveness of the model, specifically with materials that are very similar in material signature. The addition of PCA before LDA yielded reduced overfitting and improved performance with plastics that the algorithm struggled to distinguish.</p>