Dataset for 'An Automated Workflow of Deep Learning Ensemble Model with Principal Component Analysis for Enhanced Material Characterisation in Energy Dispersive Spectroscopy'

The dataset captures elemental composition information obtained via STEM–EDS spectroscopy. In STEM–EDS, each element produces characteristic peaks at specific energies corresponding to its atomic structure. The intensity of each peak is proportional to the element’s concentration in the specimen, enabling quantitative compositional analysis. This relationship allows simultaneous identification and quantification of multiple elements across the sample.Raw spectrum images in the dataset provide both qualitative and quantitative information for each pixel. Peak energies indicate elemental identity, while intensity variations reflect local concentration differences. The high spatial resolution of STEM–EDS, combined with multi-frame accumulation, ensures accurate mapping of compositional heterogeneities.For preprocessing, PCA decomposition was applied to reduce dimensionality and enhance signal-to-noise ratio, followed by Gaussian blurring and intensity distribution analysis to classify components as ‘signal’ or ‘noise’. Offline data augmentation, including colour jittering, random greyscale conversion, and Gaussian noise addition, was employed to balance the dataset. The resulting dataset comprises approximately 3,000 labelled images, partitioned into training, validation, and test subsets.This dataset provides a comprehensive resource for machine learning and computational analyses aimed at material characterisation, compositional mapping, and signal denoising in high-resolution elemental imaging studies.