Designing Waste Plastics Recycling Processes Based on Machine-Learning Thermodynamics Predictions

The chemical industry's shift towards a net-zero CO2 emissions future requires a circular economy where both mechanical and chemical recycling play a key role in reducing plastic pollution and substituting fossil-based feedstocks.[1] However, chemical recycling of Polyurethane (PUR) rigid foams, which leads to complex mixtures, often requires high-energy separation processes, undermining the overall sustainability of the recycling process.

To address this challenge, we aim to develop low-energy separation pathways for the chemical recycling of PUR rigid foams. The end-of-life PUR is processed in a catalytic process resulting in a mixture of aniline, propylene glycol, and other by-products. Distillation columns are commonly used to recover aniline, but the energy demand and the total cost of the separation are expected to be high. To address this, extraction could potentially be integrated into the process design, but identifying a suitable solvent is a challenge.

For this purpose, precise methods are required to calculate thermodynamic properties. Recently, we showed that a machine-learning approach (SPT) can precisely predict activity coefficients for a broad range of molecules.[2] Here, we expand the scope of SPT to determine the solvents' thermodynamic properties. Based on the SPT predictions, we can screen a large database of solvents suitable for aniline recovery. The resulting method yields a ranked list of solvents potentially suitable for the chemical recycling of PUR.

Our work demonstrates the conceptual design of the separation of a reaction mixture coming from PUR foam recycling enabled by machine learning. This resulting design method allows screening rapidly and accurately a very large number of solvents for which the thermodynamic properties are not experimentally available. The computational efficiency method renders the method particularly suited for future integration into the framework for integrated computer-aided molecular and process design (CAMPD).

1. Meys, et al., Science 2021. DOI: 10.1126/science.abg9853
2. Winter, et al., 2022. https://doi.org/10.48550/arXiv.2209.04135