This project aims to use the EnergyBERT(UNSW) and GPT-3(OpenAI) large language models(LLMs) to understand Light- and elevated Temperature-Induced Degradation (LeTID) in silicon solar cells. This type of degradation was identified over a decade ago and since then has >1500 publications. Despite this, there is still no consensus on the defect causing the degradation, and it remains somewhat unpredictable. The degradation has been found to affect all types of silicon, can cause up to 16% power loss and seems to be affected by variation in cell structure, process conditions or field conditions.

In this project, we will fine-tune LLMs to seek relevant LeTID data from the literature. This will be used to understand what is known from the thousands of past publications and attempt to detect patterns in data to understand what the driving factors are that cause LeTID and be able to predict its behaviour. It will also help to determine gaps in knowledge to direct ongoing experiments and will provide an opportunity to conduct important wafer, cell, or module degradation studies in the lab.

School

Photovoltaic and Renewable Energy Engineering

Research Area

Machine learning | Silicon solar cells | Degradation | Natural language processing

At SPREE, we have access to world-class research labs and the southern hemisphere's biggest supercomputer. In this project, the student will work closely with supervisor Dr. Alison Ciesla, who is Scientia and DECRA research fellow specialising in LeTID and solar module reliability, co-supervisor 3rd year PhD candidate Tong Xie who has extensive experience developing and training LLMs and have oversight support from Prof Bram Hoex. The project will be primarily software-focused to train the model. However, there will be opportunities to perform degradation studies on wafers/cells and/or modules.

The student undertaking this project will learn how to conduct a scientific literature review, including interpretation of scientific publications, especially understanding the methods of solar cell processes and critical analysis of results as required for training BERT. The student will have the opportunity to learn about how BERT and GPT-3 models are fine-tuned and perform analysis to ensure the accuracy of the outputs. The student will also learn the design of experiments and data analysis to perform required degradation studies in the lab. The outputs of this research may lead to potential journal/conference publications.