Presentation Details
| From scarce to seen: Enhancing rare defect visibility in photovoltaics using AI-generated luminescence images Gaia M.N.Javier, Brendan Wright, Victoria Zhao, Tess Rickard, Ziv Hameiri. The University of New South Wales, Sydney, Australia |
Abstract
Critical defects in solar cells are often underrepresented in luminescence image datasets, limiting the reliability of data-driven defect analysis. While prior studies have largely relied on generative adversarial networks for image synthesis, their application to luminescence imaging is often limited by training instability and limited controllability. This study, therefore, investigates alternative generative approaches for solar cell luminescence imaging to improve the visibility of rare defect patterns. A variational autoencoder, a simple diffusion model, and a conditional diffusion model were developed and evaluated using electroluminescence images of silicon solar cells. The results show that a novel combination of diffusion-based image generation with latent conditioning enables sharp synthetic images with controllable defect features. This capability enables large‑scale open-source augmentation of luminescence datasets, significantly enhancing the visibility of rare defects and thereby supporting more robust defect analysis in photovoltaic applications.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.