The most important resource of data-driven research is data. However, the amount of data available on polymeric materials is overwhelmingly small compared to other material systems. To address this data scarcity, we describe a simulation-to-real (Sim2Real) transfer learning methodology, in which a large dataset of polymer property data generated through molecular simulations is leveraged to enhance a smaller set of experimental data. Additionally, we introduce the development of RadonPy, a Python library that fully automates all-atom classical molecular dynamics simulations of polymer properties, serving as the materials informatics (MI) foundation for Sim2Real transfer learning. We also discuss the industry-academia collaboration to develop the world's largest database of computational polymer properties using the RadonPy. As of July 2024, more than 80,000 polymers have been calculated, establishing this as the world's largest such database. Using the RadonPy database as training data, we have constructed a calibration model that compensates for the gap between calculated and experimental data through Sim2Real transfer learning. Furthermore, we observed a scaling law in Sim2Real transfer learning, where the generalization performance to experimental values improves according to a power law as the amount of simulation data increases