Operating Condition-Guided Machine Learning for Solvent Selection in Polymer Solubilization

Solvent-based recycling of plastic materials is emerging as a promising strategy to leverage the recycling of complex materials, namely multipolymer products. Through polymer dissolution, complex plastic wastes can be separated into their individual components, without downgrading plastic quality. However, solvent selection remains one of the main challenges within these processes. The variety of solvents proven adequate for each polymer is still limited, and common screening approaches rely on estimations of thermodynamic parameters such as activity coefficients or Hansen Solubility Parameters. While valuable, these descriptors are obtained through separate predictive models or empirical estimations, which may introduce additional uncertainty and may not fully capture the effect of operating conditions on dissolution. Therefore, machine learning-based approaches can provide an alternative route to efficient solvent selection, identifying promising and potentially more sustainable candidates, thereby reducing the experimental burden associated with polymer-solvent testing.

This work addresses this by developing supervised machine learning models to predict polymer dissolution, while considering time, temperature, and polymer concentration as inputs, in addition to polymer and solvent’s molecular descriptors. The database was constructed using exclusively experimental data from over 70 papers, resulting in more than 900 entries, with polymers such as polyamide 6, polyamide 66, and polyvinylidene fluoride remaining largely unstudied. Gradient boosting and multilayer perceptron models were trained on this dataset. By evaluating variable importance for these models, it was possible to conclude that operating variables played an important role in distinguishing between solvents and non-solvents, but the models were also strongly impacted by the type of polymer, distinguishing halide-containing polymers, for example, and by the chemical functionalities of the solvents, namely the presence of alcohol and amide groups. This way, rooted in trends on chemical affinity and operating conditions, the models developed can be a strong tool to facilitate solvent selection for target polymers in complex materials.