Self-Improving Photosensitizer Discovery System via Bayesian Search with First-Principle Simulations

Code

Code used for photosensitizer (PS) molecular space generation, graph convolutional model training, active learning, and other analysis purposes are found in this repository.

Trained Models

The DA and DAD PS prediction models can be found here in this Google drive.

Data

Our labelled dataset of 14164 photosensitizer structures for both DA and DAD are found in data/Photosensitizers_DA.csv and data/Photosensitizers_DAD.csv respectively. All molecules are in SMILES format, and S1-T1 energy gap (ST Gap), HOMO LUMO gap (HL Gap), S1 and T1 are all in eV. Ground states were optimized with b3lyp functional and 6-31G(d) basis set. Excited-state characteristics were calculated with TD-DFT with the same level of theory using the optimized ground state geometries.

Code Authors

Jiali Li, Pengfei Cai