Analytical studies of nanoparticles (NPs) are frequently based on huge datasets derived from hyperspectral images acquired using scanning transmission electron microscopy. These large datasets require machine learning computational tools to reduce dimensionality and extract relevant information. Principal component analysis (PCA) is a commonly used procedure to reconstruct information and generate a denoised dataset; however, several open questions remain regarding the accuracy and precision of reconstructions. Here, we use experiments and simulations to test the effect of PCA processing on data obtained from AuAg alloy NPs a few nanometers wide with different compositions. This study aims to address the reliability of chemical quantification after PCA processing. Our results show that the PCA treatment mitigates the contribution of Poisson noise and leads to better quantification, indicating that denoised results may be reliable from the point of view of both uncertainty and accuracy for properly planned experiments. However, the initial data need to be of sufficient quality: these results can only be obtained if the signal-to-noise ratio of input data exceeds a minimal value to avoid the occurrence of random noise bias in the PCA reconstructions.

Scanning transmission electron microscopy is a crucial tool for nanoscience, achieving sub-nanometric spatial resolution in both image and spectroscopic studies. This generates large datasets that cannot be analyzed without computational assistance. The so-called machine learning procedures can exploit redundancies and find hidden correlations. Principal component analysis (PCA) is the most popular approach to denoise data by reducing data dimensionality and extracting meaningful information; however, there are many open questions on the accuracy of reconstructions. We have used experiments and simulations to analyze the effect of PCA on quantitative chemical analysis of binary alloy (AuAg) nanoparticles using energy-dispersive X-ray spectroscopy. Our results demonstrate that it is possible to obtain very good fidelity of chemical composition distribution when the signal-to-noise ratio exceeds a certain minimal level. Accurate denoising derives from a complex interplay between redundancy (data matrix size), counting noise, and noiseless data intensity variance (associated with sample chemical composition dispersion). We have suggested several quantitative bias estimators and noise evaluation procedures to help in the analysis and design of experiments. This work demonstrates the high potential of PCA denoising, but it also highlights the limitations and pitfalls that need to be avoided to minimize artifacts and perform reliable quantification.