Closing the DFT–Experiment Domain Gap in Redox Potential Prediction with Transfer Learning

Aqueous organic redox flow batteries (AORFBs) are a promising technology for grid-scale energy storage, using organic redox-active molecules (ORMs) as aqueous electrolytes. Redox potential ($E^\circ$) is a key property for ORM design, yet density functional theory (DFT) predictions can differ from experiment by approximately 0.2 - 0.5 V. Recent machine learning approaches have largely focused on reproducing DFT-calculated $E^\circ$, acting as rapid surrogates rather than predicting experiment directly. Here, we use transfer learning to bridge this DFT-experiment domain gap. We construct RedoxDB, containing 7,615 DFT-calculated aqueous redox potentials, and pre-train graph neural networks (GNNs) before fine-tuning on experimental data curated from the literature. Although the pre-trained models reproduce DFT accurately, they transfer poorly to experiment. Fine-tuning improves experimental prediction across all tested architectures, with our final model, RedEx, achieving $R^2=0.900$ and a mean absolute error of 0.0567 V. These results show that DFT pre-training learns chemically useful molecular representations that can be transferred to the experimental domain, providing a route to more reliable virtual screening for AORFB electrolyte development. The main code and data supporting this work are available at 'https://anonymous.4open.science/r/RedEx-85C4/README.md'.