Active Learning for Predicting Polymer/Plasticizer Phase Behaviour

Plasticisers (PLs) are small additives commonly incorporated into polymer composites to enhance processability and improve mechanical properties. Their effectiveness depends heavily on their miscibility within the polymer melt, yet isolating the influence of plasticiser properties, such as flexibility and geometry, remains challenging. This difficulty stems from the time consuming nature of experimental work and also from the presence of impurities and inconsistencies that often arise during synthesis and testing. Atomistic simulations face similar difficulties as phase separation can occur over microsecond timescales, which can be computationally expensive. In this work, we use a coarse-grained bead-and-spring model to screen plasticisers of varying flexibilities and geometries to build a pool-based active learning procedure which characterizes their design space and its effect on polymer/plasticiser miscibility. We perform an active learning cycle with a random forest model and an uncertainty/random hybrid query strategy to determine the miscibility behaviour of queried molecules. This is evaluated through computationally expensive, coarse-grained polymer/plasticiser simulations of a cis-(1,4)-polyisoprene melt filled with small hydrocarbon additives of varying sizes and rigidities. Through this, we are able to efficiently improve model performance in order to make predictions on the entire PL design space. Such findings enable us to determine a new set of general plasticiser design rules, suitable for non-polar molecules, which expands on our previous work. To further prove this, we perform atomistic simulations of polyisoprene/plasticiser systems which are approximately back-mapped from their coarse-grained equivalents. Our findings indicate that the polyisoprene/plasticiser phase behaviour, observed using the coarse-grained model for PLs in the absence of side chains, is valid.