Data-driven Materials Informatics for Optoelectronics: From Natural Language Processing to Predictive Modelling of TADF Molecules

This thesis addresses the development and application of data-driven approaches to materials informatics for optoelectronics, with a focus on thermally-activated delayed fluorescence (TADF). Chapter 1 provides an introduction to the background and recent progress in data-driven methods in materials sciences and thermally-activated delayed fluorescence. Chapter 2 reviews the natural language processing techniques and language modelling methods that were used throughout the thesis. Chapter 3 demonstrates a pipeline for the extraction of four organic TADF molecule property data from the literature, namely, maximum emission wavelength (𝜆EM ), photolu- minescence quantum yield (PLQY), singlet-triplet energy splitting (Δ𝐸ST ), and delayed life- time (𝜏D ). The pipeline affords a database of 25,482 data records with a collective precision of 82%. Chapter 4 describes a cost-efficient approach to pre-training “optoelectronics-aware” language models via domain-adaptative pre-training (DAPT). Three language models, OE- ALBERT, OE-BERT, and OE-RoBERTa, are produced using this approach. They are also fine-tuned to perform tasks of text-classification, question-answering, and text embedding. Chapter 5 details an end-to-end workflow that produces a data-driven predictor for the PL wavelengths of organic TADF molecules using molecular SMILES strings as its input. The workflow utilizes techniques developed in Chapter 3 and 4 to collect training data. The predictor achieves accurate PL wavelength estimation on an out-of-sample test set with a mean absolute error of 0.13 eV. Chapter 6 concludes the work and discusses potential directions for future research.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2025

Author dc:creator

• Huang, Dingyun

Advisor dc:contributor.advisor

• Cole, Jacqueline

• deep learning
• language model
• machine learning
• organic light-emitting diode
• text-mining
• thermally-activated delayed fluorescence

Licence

• https://www.repository.cam.ac.uk/bitstreams/8019e4bf-8584-4343-a09b-1c0448ea9a41/download
• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.125133

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/395706