Transfer learning of GW Bethe--Salpeter Equation excitation energies

Key Points

• The research aims to improve the prediction of excitation energies in electronic-structure calculations using transfer learning methods.
• Developed a transfer learning framework utilizing low-fidelity data like DFT results.
• Applied machine learning techniques to predict high-fidelity excitation energies from sparse data sets.
• Conducted comparisons between predictions and established benchmarks.
• Successfully increased accuracy of excitation energy predictions compared to traditional methods.
• Showed significant improvements in prediction reliability with limited high-fidelity data.
• Demonstrated that transfer learning can mitigate the data scarcity issue in electronic-structure modeling.