Closing the loop for the discovery of novel catalysts and mechanisms for depolymerization of functional materials
Abstract
In general, there are three types of plastic recycling methods: pyrolytic, physical andchemical. The vast majority of plastics can be pyrolytically recycled, i.e. converted into fuels, butthis still results in carbon emissions. Physical recycling of plastics, e.g., polyethylene andpolyethylene terephthalate (PET), often involves degradation and downcycling. Chemicalrecycling of commodity plastics into chemical feedstocks has dual benefits. Rather than needingto extract oil from ever-depleting reserves, valuable feedstocks generated from chemical recyclingof commodity plastics will enable a closed carbon cycle, effectively upcycling the polymers andkeeping the carbon dioxide levels in the troposphere stable.Herein, we propose to develop new strategies for chemical recycling of polypropylene,because it is one of the polymers most difficult to recycle chemically and second most abundantplastic. Its inherent chemical stability makes selective chemical recycling challenging anddemands for a data-driven closed-loop optimization approach conducted in self-driving labs.Wewill largely rely on photoredox-based oxidative fragmentation of polypropylene as the energyresulting from photoexcitation of the catalyst enables the selective cleavage of strong chemicalbonds. Materials research and discovery has to move past the traditional and time-consumingmodel of serendipity and trial-and-error improvements. Automated laboratories accelerate thepace of laboratory work significantly. Using modern AI-based decision-making, these laboratoriescan be made truly self-driving, moving towards experimenter-selected destinations, i.e.materials with desired properties, without human steering.Bayesian optimization is considered the gold standard for improving quantities that areexpensive to evaluate. Still, standard methods require that inputs can be ordered, as is the case,e.g., for temperature or concentration. These methods do not apply to categorical optimization,where the problem inputs have no well-defined or known ordering. Categorical variables include,for example, the chemical identity of reagents. Consequently, Bayesian optimization had, untilrecently, somewhat limited applications in chemistry.In this project, we will employ closed-loop optimization to find optimum catalysts andreaction conditions. High-throughput experimentation guided by our ML-based optimizationsoftware platform, ChemOS, will enable more efficient discovery of new catalysts by decreasingthe number of necessary experiments. Furthermore, experimental optimization will go hand inhand with the development of better representations to be employed within ChemOS. Moreover,implementing the closed-loop discovery workflow in experimental platforms necessitates handlinginfeasible experiments, i.e. experiments that do not result in a foreseeable outcome, which is oneof the current drawbacks of the algorithms implemented in ChemOS. The outcome will be a robustclosed-loop discovery workflow for the development of new catalysts and chemical reactions anddisrupt the way discovery of novel catalytic protocols is currently carried out
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 15, 2021
- Source ID
- N000142112137
Entities
People
- Aln Aspuru-guzik
Organizations
- Office of Naval Research
- United States Navy
- University of Toronto