Machine Learning Regression-Enhanced Property Prediction of Brush Particle Solids

The development of advanced hybrid materials such as polymer-grafted nanoparticles (aka "brush" or "hairy" particles) hinges on the ability to accurately predict suitable compositions to realize desired functionalities. Such systems are characterized by expensive synthesis, sparse data, and large parameter spaces that often render physics-based modeling or experimental exploration of structure-property relationships expensive and impractical. Regression models provide a versatile data-driven alternative to establish empirical structure-property relations. However, the implications of sparse data availability and method of training on the predictive capability of regression models outside of the respective range of training data is poorly understood. This work presents a comparative analysis of widely used linear and nonlinear models versus machine learning regression models to evaluate their ability to predict the elastic modulus of brush particle solids formed by poly-(methyl methacrylate) (PMMA) and polystyrene (PS) tethered silica nanoparticles with systematic variation of the degree of polymerization, dispersity, and grafting density. The methods investigated include k-nearest neighbors, multilayer perceptron, and XGBoost regression. The performance of each model is assessed on a consistent set of training and testing data. Within the experimental certainty, the k-nearest neighbors regressor and XGBoost were found to provide the most reliable property prediction. Combining data sets of disparate polymer systems (PMMA/PS) was found to increase the accuracy of prediction and to enable models to forecast previously reported unexpected behavior outside of the range of the original training data. This supports the use of trained regression models as capable exploratory tools to reveal structure-property relations for systems that are controlled by multiple parameters with unknown interactions.