From Generation to Validation: Ensuring the Scientific Utility of Synthetic Metallographic Microstructures

Daria M. Tomecka¹, S. Khalatabad², R. Jaeken¹, B. Lis¹, L.A.I. Kestens², S. Cottenier², M. Sluydts¹

¹ePotentia, Frans van Dijckstraat 59, B-2100 Deurne Antwerpen, Belgium,

²Ghent University, Department of Electromechanical, Systems and Metal Engineering, Tech Lane Ghent Science Park - Campus A, Technologiepark 131, B-9052 Gent, Belgium

daria.tomecka@epotentia.com

The growing adoption of AI in materials characterization demands increasingly large, diverse, and unbiased datasets. However, acquiring comprehensive steel microstructure data through electron microscopy remains resource-intensive and often leads to incomplete or imbalanced coverage of process conditions. While generative AI offers a promising solution through synthetic data creation, validating these artificial microstructures raises fundamental questions about authenticity and scientific utility.

Within the AID4GREENEST project [1], we use the Ultra High Carbon Steel Database (UHCSDB) [2] to explore the entire synthetic data pipeline - from generation to validation. Synthetic micrographs are produced using a phase- and attribute-aware sampling strategy based on real data distributions. We evaluate how well these microstructures reflect realistic phase compositions, morphologies, and variations driven by process parameters such as annealing temperature, time, cooling method, and magnification. Our study critically examines standard AI-based image similarity metrics in both pixel and Fourier space, questioning their capacity to preserve physical and structural properties. These computational validations are compared with expert metallurgists’ assessments to examine whether current approaches align with domain expertise in distinguishing real from synthetic microstructures.

This comprehensive analysis seeks to establish robust criteria for evaluating synthetic microstructures, ensuring that generated data not only resembles real steel samples visually, but also retains scientific utility for downstream materials analysis.

References

[1] AID4GREENEST. (2025). [Online]. Available: https://www.aid4greenest.eu

[2] DeCost, B. L., Hecht, M. D., Francis, T., Webler, B. A., Picard, Y. N., & Holm, E. A. (2017). UHCSDB: UltraHigh Carbon Steel Micrograph DataBase: Tools for Exploring Large Heterogeneous Microstructure Datasets. Integrating Materials and Manufacturing Innovation, 6(2), 197–205.

https://doi.org/10.1007/s40192-017-0097-0

Figure 1 (see pdf). Mapping the microstructure space: Each point represents a single micrograph. On the right, a zoomed-in micrograph is shown; can you determine if it is real or synthetic?

Keywords: Steel microstructures; SEM; Synthetic data; Deep learning; FFT; Image similarity metrics; Metallographic validation; Process control.