Predicting Solute Diffusivity and Transport Kinetics in Polymers Using Quantile Random Forests

Abstract

Additives and contaminants in polymer-based medical devices may leach into patients, posing a potential health risk. Physics-based mass transport models can estimate the leaching kinetics, but they require upper-bound estimates of solute diffusivity $D$ in the polymer. Experiments to measure $D$ can be costly and time-consuming. Alternatives to estimate $D$ exist, but they suffer from several drawbacks, such as requiring experimental data to calibrate or specialized knowledge to apply, being limited to certain polymers, or being too time-consuming given the plethora of polymer/solute combinations in devices. Here, we leverage a large database of diffusivity measurements and apply a machine learning method—quantile random forests (QRF)—to predict bounds on $D$ for arbitrary polymer/solute combinations, using only the solute structure and readily available polymer properties (glass transition temperature $T_g$ and density). The most influential factors for determining $D$ are these polymer properties and several descriptors related to solute size (e.g., molecular weight $M_w$), structure, and interactions. Note that application of the model is limited to the applicability domain defined herein and polymers with relatively low fractional-free-volume. We demonstrate the ability of the model to predict $D$ and diffusion-limited transport kinetics, where it compares favorably to other available methods while also overcoming the aforementioned drawbacks.