PolyMon: A Unified Framework for Polymer Property Prediction With Imperfect Simulators

Accurate prediction of polymer properties is essential for polymer design, but remains challenging due to data scarcity and the fact that the more plentiful data from physics-based simulations may not fully capture polymer behaviour. How to best exploit such simulator-derived data alongside limited real-world measurements remains an open question. We present PolyMon, a unified framework that integrates diverse polymer representations, ML models, and training strategies for leveraging imperfect simulators: multi-fidelity learning, $\Delta$-learning over physics-based estimators, active learning with MD simulations in the loop, and ensemble learning. Using five polymer properties as benchmarks, we systematically evaluate how representations, models, and training strategies affect predictive performance. We find that (1) learning the residual over a simulator-pretrained model improves experimental density prediction; (2) $\Delta$-learning provides a powerful means of incorporating knowledge from physical estimators and other property datasets; and (3) uncertainty-guided active learning with MD labelling in the loop is more data-efficient than random acquisition. Overall, PolyMon provides a comprehensive and extensible foundation for benchmarking and advancing ML-based polymer property prediction. The code is available at https://anonymous.4open.science/r/polymon.