Plasmon-Driven Degradation of Organic Contaminants:Accelerating Optimization through Bayesian Experimental Design

Abstract

The accumulation of plastic waste is one of the most urgent environmental challenges of the twenty-first century. Global plastic production has increased from around 2 million tonnes in 1950 to over 400 million tonnes per year, only 9% is recycled into equivalent-quality products.1 In Europe, strict food-contact regulations make recycling more difficult due to contaminants absorbed during the service life of a product (plasticizers, residual solvents, and volatile organic compounds), which make post-consumer plastics unsuitable for food-grade reuse without thorough decontamination.2 Traditional approaches such as supercritical CO2 extraction and steam stripping consume high energy inputs and can damage the polymer matrix.3 Photocatalysis represents a viable alternative, as it uses light as the primary energy input to generate reactive oxygen species on semiconductor and plasmonic surfaces under ambient temperature and pressure, in line with green chemistry and circular economy principles. Composite photocatalysts combining graphitic carbon nitride (g-C3N4), a visible-lightactive polymeric semiconductor, with plasmonic gold nanoparticles (Au@g-C3N4) are particularly attractive: where the g-C3N4 acts as the semiconductor host with the main photocatalytic engine stimulated by UV-blue light, while the gold nanoparticles act as plasmonic light-harvesting antennae to extend the spectral response of the system deep into the visible range through localized surface plasmon resonance, and the resulting heterojunction suppresses charge recombination.4,5 However, optimizing such systems is challenging because the multi-dimensional parameter space (spanning catalyst composition, synthesis conditions, and reaction variables) cannot be efficiently explored by traditional one-variable-at-a-time experimentation. Bayesian Optimization (BO), which iteratively builds a probabilistic surrogate model and selects experiments that balance exploration of uncertain regions with exploitation of promising ones, is well suited to this challenge. The present work is conducted within the FFG-funded LightAIClean project, a collaborative initiative developing AI-driven, light-based decontamination technologies to enable food-contact-compliant bottle-to-bottle recycling. In this thesis, the photocatalytic degradation of Methylene Blue, as a model organic contaminant, by Au@g-C3N4 composites is systematically characterized and optimized, and an AI-guided experimental workflow is validated. Baseline performance of individual components is established for AuNPs (~13 nm) with 525 nm LED illumination and for g-C3N4 with 450 nm illumination, along with requisite control experiments. Performance of Au@g-C3N4 composites is mapped by varying four key parameters (solution pH 5-11, catalyst dosage 0.1-2.0 g/L, gold loading 0.5-5.0 wt%, and preparation stirring time 0-23 h), with the irradiation time and illumination wavelength additionally varied as optimization parameters in the subsequent AIguided round. These seed data feed a Bayesian Optimization model to suggest targeted experiments to identify global performance optima efficiently. Performance is measured in terms of pseudo-first-order rate constants, degradation efficiencies, and apparent quantum yields. The validation of this transferable methodology will enable it to be extended to industrially relevant contaminants and accelerate the development of practical decontamination technologies for circular plastic recycling.