Yug Shah
Research Mentor: Majdi Radaideh
Mentor Department: Nuclear Engineering and Radiological Sciences, Engineering
Author(s): Yug Shah, Majdi Radaideh, Mohammed Radaideh
Session: Session 6 (3:00 PM – 3:50 PM)
Presentation Type: Poster 18
Abstract
Generative AI has shown potential as a tool to enhance public engagement and energy literacy, particularly around low-carbon technologies such as nuclear power, but current mainstream AI models have some limitations when dealing with these ideas. In our prior study, we assessed 20 models, including DALL-E and Craiyon, and found that even though they could generate images for general prompts, they often produced inaccurate technical depictions, reinforced gender biases in the energy sector, and poorly represented Indigenous landscapes. To address these limitations, this project employs a domain-specific approach by fine-tuning the open-source text-to-image generative AI Stable Diffusion models (V1.5, 3.5) on nuclear energy images, which were derived by parsing through thousands of scholarly articles. The methodology involves tweaking the model to improve technical accuracy and inclusivity in the generated images. The expected results are that the fine-tuned model will outperform mainstream text-to-image AI tools in technical accuracy, cultural representation, and reducing gender bias in subjects related to nuclear energy. By producing scientifically and socially accurate images, this work aims to increase public engagement in clean energy discussions. The results of this research could demonstrate the need for specialized AI tools to improve public understanding in critical but complex and somewhat controversial fields.
