Machine learning climbs the Jacob's Ladder of optoelectronic properties.

The use of machine learning (ML) as a powerful tool for the prediction of optoelectronic properties is still hampered by the inadequate level of the calculated training datasets, which are almost exclusively obtained within the independent-particle approximation (IPA). Drawing on Perdew's Jacob's ladder analogy in density functional theory, we demonstrate how ML can ascend from the IPA to the random phase approximation (RPA), figuratively climbing the second rung. We show that as few as 300 RPA calculations suffice to fine-tune a graph attention network initially trained on 10,000 IPA calculations. Its prediction accuracy approaches that of a network directly trained on our large database of around 6000 RPA spectra. Our results highlight how transfer learning even with a small amount of high-fidelity data significantly improves predicted optical properties. Moreover, by retraining on RPA data from materials with smaller unit cells, the model generalizes effectively to larger unit cells, demonstrating broad scalability.