Human-Guided vs. Push-Button Machine Learning for Hidden-Phase Raman Mapping

Raman mapping routinely acquires thousands of spectra, yet fluorescence, weak signal and peak overlap can mask key phase information. We contrast two open-source workflows – an interactive, human-guided pipeline and a fully automated machine-learning routine – to reveal hidden structural differences in complex (geo)materials.

The case study is a cross-section of a prior analysed oxide-rich corrosion layer on a ferritic model alloy. Corrosion control demands separating protective from non-protective oxides e.g., magnetite, chromite and mixed spinels. In the human-guided routine, an analyst applies dimensionality reduction, chooses a clustering method, inspects cluster quality and, where warranted, performs pseudo-Voigt peak-position analysis before re-clustering. This loop resolves all the chemically distinct populations and quantifies magnetite – chromite band shifts of up to 7 cm^-1, revealing mixed spinels that standard Raman maps miss.

The automated path eliminates human decisions: a convolutional autoencoder compresses spectra to latent features and an information-criterion optimiser determines cluster numbers. While faster, the automated routine tends to merge subtle spinel variants and is more sensitive to noise. The present study dissects trade-offs in interpretability, reproducibility and library dependence, showing when expert oversight adds value.

These workflows turn phase-tagged Raman maps into quantitative micro-atlases that can be generated on notebooks using scikit-learn^[1] and TensorFlow^[2]. The presentation shows how both workflows deliver large-scale, detailed insight into chemically similar phases with differing functions.

References

[1] Pedregosa F. et al. (2011) J. Mach. Learn. Res. 12, 2825 – 2830.

[2] Abadi M. et al. (2016) Proc. 12^th USENIX Symp. (OSDI 16), 265 – 283.