Transfer learning for predicting reorganization energy

Reorganization energy, a crucial factor in quantum chemistry, plays an integral role in the research of optoelectronic and electrical devices such as organic field-effect transistors and organic light-emitting diodes. Currently, the calculation of reorganization energy predominantly depends on density functional theory, which poses significant computational costs. Traditional machine learning, commonly employed to predict reorganization energy necessitates a substantial volume of data for model training, a process that can be time-consuming due to difficulties in collecting sufficient data. This study proposes a novel solution by employing transfer learning. This method involves the pre-trained model using publicly accessible large-scale data on frontier molecular orbital energy, and then transfers this model to predict the reorganization energy. The experimental results indicate that the proposed method requires approximately 28% less data compared to traditional machine learning while maintaining equivalent accuracy. Moreover, this method achieved a median error of less than 0.02 eV when predicting low reorganization energy. Consequently, the need for reorganization energy data in model construction is significantly reduced, thus facilitating research into reorganization energy and expediting research into the optical properties of materials.