Flexible electronics are becoming the next generation of devices due to their advantages, such as mechanical flexibility, eco-friendliness, large-area applicability, and scalability for mass production. However, solution-based manufacturing processes are prone to defects like discontinuities and local smudging, which can significantly degrade both device quality and yield. To tackle these challenges, rapid and accurate defect classification is crucial for real-time diagnosis during manufacturing. This study investigates the impact of data scale and key training hyperparameters on the performance of deep learning–based defect diagnosis models, using a dataset of conductive pattern defects in flexible electronics. We specifically examine how the number of training images affects model accuracy and generalization, and we analyze how adjustments to hyperparameters—such as L2 regularization and dropout—influence model performance in data-limited scenarios. Our findings offer insights into optimal training strategies tailored to different data scales and learning constraints, providing practical guidelines for designing and developing AI-based defect diagnosis models for flexible electronic devices.
