Synthetic Data Generation for Pharmaceutical Based Microplastic Detection in Water Systems

Microplastic pollution is a major threat to the environment and human health, but existingdetection methods are far too expensive and slow to enable large-scale monitoring. This projectbuilt a computer simulation that models how microplastics can be detected using Nile Redfluorescent dye and photodiode sensors, creating synthetic datasets to train machine learningalgorithms while still behaving like actual physical systems by accounting for how differentplastics fluoresce under UV light, how dye molecules bind to plastic surfaces through Langmuiradsorption, natural particle size distributions, and the noise inherent in real sensors. Laboratorytests with polystyrene and polyethylene microplastics validated the model, which matchedexperimental results with just 3.2% average error across all conditions tested. Sobol sensitivityanalysis showed that particle size drove around 46% of the variation in fluorescence signalswhile plastic type contributed to about 31%, demonstrating that measuring both particle size andbrightness gives much better classification than brightness alone. A test Machine learningmodels trained on 15,000 simulated samples correctly identified plastic types 89.7% of the timeand predicted microplastic concentrations with an R² of 0.94. This work connects expensive labexperiments with practical algorithm development, letting researchers test thousands of differentconfigurations, refine sensor designs, and train detection algorithms on computers beforespending money on actual hardware, which provides a starting point for building low-cost,automated microplastic detection systems that could be installed at water treatment plants andmonitoring stations around the world.