Combining Flory-Huggins Theory and Machine Learning for Improved Polymer Solution Phase Behavior Predictions

Polymer solution macrophase separation predicted by Flory-Huggins theory often shows disagreement with experiment. Improvements rely on empirical parameters fit to modified Flory-Huggins equations that require extensive experimental data, complex optimization, and are ultimately chemistry-dependent (i.e., useful for a small range of polymer solutions). Recently, machine learning (ML) models trained on experimental data have shown predictability of the phase boundary to within 1-3^oC across a broad range of polymer-solvent systems and extrapolation to new systems with as little as 20 additional data points. However, the physics underlying these ML predictions are obscured, which hinders human interpretability and trust. In this work, we build a framework combining Flory-Huggins theory with ML models to (i) improve predictability and generalization with less experimental data, and (ii) provide interpretability. To achieve this, we estimate theoretical, molecular weight dependent parameters using neural networks. Using polystyrene-cyclohexane as a focus, we compare the ability of these hybrid theory-ML models relative to current theory and traditional ML to predict macrophase separation, and interpolate to new molecular weights as a function of training data set size. Our framework showcases the benefits of physics-interpretable ML models for polymers.