The depolymerization of polyethylene terephthalate (PET) through efficient chemical recycling remains a central challenge in plastic waste valorization, in part because the catalyst landscape is vast and sparsely explored. Here, we present an artificial intelligence (AI)-driven discovery framework that integrates Bayesian optimization (BO), large language models (LLMs), and high-throughput robotics to accelerate the search for Lewis acid-base catalysts for PET glycolysis. Starting from a literature-guided baseline, BO used LLM-derived semantic embeddings of chemical knowledge to navigate a high-dimensional space of 11,160 candidate pairs, identifying promising candidates beyond the initial state of the art. The LLM then analyzed the experimental results to generate interpretable, data-driven hypotheses that guided further experiments and enabled inductive, human-led extrapolation beyond the predefined search space. This workflow yielded a zinc pivalate/N,N'-diethylethylenediamine catalyst delivering 95% bis(2-hydroxyethyl) terephthalate (BHET) yield in 20 min, with robust performance upon scale-up and on postconsumer PET. Mechanistic analysis supports a synergistic dual-site activation mode and informs transferable design principles. All experiments were executed on a fully autonomous AI-Chemist platform with automated reaction setup and nuclear magnetic resonance (NMR) spectroscopic analysis. Together, these results show how automation-AI-human collaboration can progress from optimization to out-of-sample discovery in large, underexplored chemical spaces.
Yu et al. (2026) studied this question.