High-throughput virtual screening of molecules for photon conversion
File(s)
Author(s)
Verma, Shomik
Type
Thesis
Abstract
Photovoltaics (PV) have emerged as a prominent technology to generate electricity from sunlight. However, traditional single-junction PV cells such as silicon, thin film PV, and perovskites suffer from an inherent efficiency limit of 33.7%. This is primarily due to two loss mechanisms: sub-bandgap losses, where photons with energy below the bandgap of the PV cell cannot be utilized, and thermalization losses, where photons with excess energy above the bandgap lose their excess en- ergy to heat. Photon conversion materials can help overcome the detailed-balance limit by converting wavelengths of light into energies the solar cell can efficiently absorb. The two common mechanisms for photon conversion are triplet-triplet an- nihilation (TTA) up-conversion and singlet fission (SF) down-conversion. Several molecules have been shown to exhibit TTA or SF, but there could be cheaper or less complex molecules previously overlooked that would be suitable. To identify such chromophores, high-throughput virtual screening (HTVS) of large databases is re- quired.
Both TTA and SF involve the singlet and triplet excited states of molecules, so knowing these excited state energies is critical. The central issue to HTVS is that limited excited state databases exist, and computational techniques for calculating excited state energies are time-consuming. This thesis aims to solve this issue with various approaches. First, triplet excited state energies are predicted with a machine learning (ML) model trained on a dataset of TD-DFT energies generated with active learning (AL) to ensure the training set size is optimized. While directly predicting energies with ML is fast, there are issues with accuracy and training time. The sec- ond approach calibrates a high-throughput computational chemistry method called xTB-sTDA against TD-DFT with ML. This ensures both high accuracy and low com- putation time. Finally, the third approach applies xTB-ML to a large dataset, using AL to actively suggest candidate chromophores for photon conversion.
Both TTA and SF involve the singlet and triplet excited states of molecules, so knowing these excited state energies is critical. The central issue to HTVS is that limited excited state databases exist, and computational techniques for calculating excited state energies are time-consuming. This thesis aims to solve this issue with various approaches. First, triplet excited state energies are predicted with a machine learning (ML) model trained on a dataset of TD-DFT energies generated with active learning (AL) to ensure the training set size is optimized. While directly predicting energies with ML is fast, there are issues with accuracy and training time. The sec- ond approach calibrates a high-throughput computational chemistry method called xTB-sTDA against TD-DFT with ML. This ensures both high accuracy and low com- putation time. Finally, the third approach applies xTB-ML to a large dataset, using AL to actively suggest candidate chromophores for photon conversion.
Version
Open Access
Date Issued
2021-09
Date Awarded
2021-11
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Walsh, Aron
Hoye, Robert
Sponsor
Marshall Scholarships
Publisher Department
Materials
Publisher Institution
Imperial College London
Qualification Level
Masters
Qualification Name
Master of Philosophy (MPhil)