Repository to the paper "Transfer learning of GW Bethe-Salpeter Equation excitation energies" (10.1039/D5SC09780K). We show how DFT and TDDFT data can be leveraged to 1) reduce test errors of GNN-based predictions of qsGW quasiparticle energies and GW-BSE excitation energies and 2) reduce the required amount of costly qsGW and GW-BSE data needed for doing so.
The filenames of model checkpoints of pretrained models in the checkpoints/ directory follow the pattern:
pre-prop-npre-M-nfeat-rcut-nbas-nmp-lmax
For instance, prehomo1M128502532 denotes a model pretrained on 1,000,000 (1M) HOMO energies (prop) with 128 feature channels (nfeat), using a cutoff distance of 5.0 Å (rcut with the decimal point left out in the filenames), 25 radial basis functions (nbas), 3 message-passing layers (nmp) and lmax of 2. The filenames of pretrained and finetuned models follow the pattern:
prop-npre-M-nfeat-rcut-nbas-nmp-lmax
For instance, homo1M128502532 denotes a model pretrained and finetuned on HOMO energies. The filenames of the baseline models without pretraining, so 0 pretraining samples, and only finetuning follow the pattern:
prop-0M-nfeat-rcut-nbas-nmp-lmax
The ViSNet model used in this work was taken from the official repository github.com/microsoft/AI2BMD