Calculating or measuring phonon properties are orders of magnitude more resource-intensive than electron properties. We demonstrate that Transfer Learning can leverage knowledge learned from electron property data to significantly improve phonon property prediction, even if the volume of phonon data is merely 10% of electron data. This great improvement is achieved despite known inaccuracy in electron property data (e.g., DFT under-predicted bandgap) — Transfer Learning works as long as source data encode some, even if imprecise, structure-property relationships. This has important implications as one can potentially use not-so-accurate high-throughput computational data to help training machine learning models using accurate but very limited experimental data. Read our Science Advances paper: https://advances.sciencemag.org/content/6/45/eabd1356