Evaluating Synthetic Images as Effective Substitutes for Experimental Data in Surface Roughness Classification

A research team led by Chi Ho Wong from The Hong Kong Polytechnic University has published a groundbreaking study demonstrating that synthetic images generated by AI can serve as viable substitutes for expensive experimental data in industrial applications. Their paper, "Evaluating Synthetic Images as Effective Substitutes for Experimental Data in Surface Roughness Classification," shows that using Stable Diffusion XL to create synthetic images of ceramic surfaces yields classification accuracies comparable to those achieved with exclusively real, experimentally acquired images. This addresses a major bottleneck in deploying AI for materials science: the need for large, labeled datasets and costly high-resolution imaging equipment.

The researchers systematically tested their approach by augmenting authentic datasets with AI-generated images and evaluating performance across various training configurations. They identified specific hyperparameter settings (epoch count, batch size, and learning rate) that preserve model performance while significantly reducing data requirements. The study confirms that synthetic images effectively reproduce the structural features necessary for accurate surface roughness classification, a critical quality metric for hard coatings used in demanding mechanical applications.

This work provides a concrete methodology for improving data efficiency in materials-image analysis. By leveraging generative AI, engineers and researchers can lower experimental costs, accelerate model development cycles, and expand the applicability of computer vision to areas where data collection has traditionally been prohibitive. The approach demonstrates robustness and offers a scalable solution for industrial AI workflows where reliability and cost-effectiveness are paramount.