Predicting Phase Behavior of Linear Polymers in Solution Using Machine Learning

Abstract

The phase behavior of polymers in solution is crucial to many applications in polymer processing, synthesis, self-assembly, and purification. Quantitative prediction of polymer solubility space for an arbitrary polymer–solvent pair and across a large composition range is challenging. Qualitative agreement is provided by many current theoretical models, but only a portion of the phase space is quantitatively predicted. Here, we utilize a curated database for binary polymer solutions comprised of 21 linear polymers, 61 solvents, and 97 unique polymer–solvent combinations (6524 cloud point temperatures) to construct phase diagrams from machine learning predictions. A generalizable feature vector is developed that includes component descriptors concatenated with state variables and an experimental data descriptor (phase direction). The impact of several types of descriptors (Morgan fingerprints, molecular descriptors, and Hansen solubility parameters) to encode polymer–solvent interactions is assessed. Hansen solubility parameters are also introduced as a means to understand the general breadth of the linear polymer–solvent space as well as the density and distribution of curated data. Two common regression algorithms (XGBoost and neural networks) establish the generality of the descriptors; provide a root mean squared error (RMSE) within 3 °C for predicted cloud points in the test set; and offer excellent agreement with upper and lower critical solubility curves, isopleths, and closed-loop phase behavior by a single model. The ability to extrapolate to polymers that are very dissimilar from the curated data is poor, but with as little as 20 cloud points or a single phase boundary, RMSE error of predictions are within 5 °C. This implies that the current model captures aspects of the underlying physics and can readily exploit correlations to reduce required data for additional polymer–solvent pairs. Finally, the model and data are accessible via the Polymer Property Predictor and Database (3PDb).

Emerging Trends in Machine Learning: A Polymer Perspective

“…While there have been several recent studies focusing on developing active learning techniques for polymers using premeasured data sets, theory or simulations, − here we focus on experimentally realized autonomous platforms. These studies either directly or indirectly address key challenges in applying ML to polymers as outlined in Table .…”

The Rise of Machine Learning in Polymer Discovery

“…[99] Ethier et al used gradient boosting to do quantitative prediction for the phase behavior of polymers in solution. [100] Figure 25. Methods for predicting the spectra of conjugated polymers.…”

Determining ion activity coefficients in ion-exchange membranes with machine learning and molecular dynamics

“…Specifically, in polymer informatics, these methods have found successful applications in predicting key properties such as glass transition temperature, 23,24 gas permeability, [25][26][27] and phase diagrams. 28 Increase in computational power has boosted the growth of physics-based tools like molecular dynamics (MD). For example, MD simulation has been used to study polymer phase behavior 29,30 and estimate properties such as thermal conductivity, 31,32 water diffusion constant, 14 and diffusion coefficient.…”

“…In this work, three strategies are combined to produce input features that integrate molecular structure information, system conditions, and molecular dynamics calculations. First, Morgan Fingerprints (MF), 45 which have found tremendous applications in ML, 23,27,28 was selected to represent the molecular structure of the Cl -…”