Sim2Real Multitask Learning for Predicting Polymer and Small-Molecule Miscibility

The miscibility of polymers and small molecules is a critical property in applications such as plastic recycling, polymer synthesis, and purification. This miscibility is described by the free energy of mixing, derived from Flory-Huggins interaction parameters. Here, we introduces a machine learning approach to predict the Flory-Huggins interaction parameters, aiming to predict the miscibility of polymers and small molecules. The significant challenge is the insufficient and biased data due to the high cost of experiments. This limitation results in low prediction accuracy for structures in extrapolation region, where the structures in that region are dissimilar to those in the training data. To address this problem, we expanded the chemical space coverage by generating the Flory-Huggins interaction parameter data through high-throughput COSMO-RS simulation based on DFT calculations. We trained the experimentally observed and simulated data simultaneously through multitask learning. This successfully enabled predictions for the extrapolation region beyond the chemical space of the training data. Our results surpassed the accuracy of a traditional method based on the Hansen solubility parameter (HSP). Moreover, we observed a scaling law, that is, the accuracy is improved with an increased number of the COSMO-RS simulation data. We anticipate further accuracy improvement with an increased simulation dataset. Our method based on the multitask learning with high-throughput simulation data is not only useful for predicting miscibility, but also has the potential to be a solution to the small data problem that is a challenge in materials informatics.