Publications by authors named "Leonardo Rotondi"
Article Synopsis
- A deep learning model predicts excited singlet-triplet gaps in molecules, achieving a mean absolute error of about 20 meV, which helps identify potential inverted singlet-triplet candidates.
- The model utilizes advanced spherical message passing graph neural networks to analyze molecular structures from Cartesian coordinates and atomic numbers, trained on a large dataset of around 40,000 diverse geometries.
- Performance varies based on the quality of the geometries used, showing good results with DFT and GFN2-xTB geometries but with notable declines when using lower-quality models, while also producing promising predictions for identifying IST candidates from both DFT and experimental structural data.
