EPFL scientists have created an AI-based approach to improve chemical analysis of nanomaterials, overcoming problems caused by noisy data and mixed signals, according to a study published in Nano Letters.
Image Credit: Yurchanka Siarhei/Shutterstock.com
A single unit of a chemical substance or material sized between 1 and 100 nanometers is referred to as “nanomaterials”; a nanometer is a billionth of a meter. These comprise unusual substances, including silver nanoparticles (used as antimicrobials), carbon nanotubes, nanoporous materials, and a variety of catalysts that effectively accelerate chemical processes.
The precise chemical makeup of nanomaterials is crucial as they are now employed in various industries, including electronics and medical. This is difficult, though, as conventional techniques for examining nanomaterials are frequently vulnerable to poor signal-to-noise ratios.
For instance, scanning transmission electron microscopy in conjunction with energy-dispersive X-ray spectroscopy (EDX) is one widely used technique. This method offers thorough maps of the locations of many elements in a sample; nevertheless, it frequently yields noisy data, particularly for small objects, and mixed signals when different materials overlap, which makes chemical analysis more challenging.
Noisy data are often “cleaned up” using various techniques, ranging from simple spatial filtering to more advanced machine learning algorithms such as principal component analysis, which separate the signals from the noise, but they also have downsides. For example, they may introduce inaccuracies or difficulty in discerning between chemical signals that are quite similar.
Hui Chen, Duncan Alexander, and Cécile Hébert, three EPFL scientists, have developed a machine learning-based method known as PSNMF ("non-negative matrix factorization-based pan-sharpening") that improves the clarity and accuracy of EDX data, making it easier to identify and quantify various chemical elements in nanomaterials.
The team began by using a unique feature of their data called “Poisson noise”. This sort of noise develops when X-ray photons are detected at random. When an electron beam strikes a sample, it emits X-ray photons, but the amount detected fluctuates randomly each time, resulting in a chaotic, grainy pattern known as Poisson noise.
The researchers combined data from nearby pixels to increase the signal-to-noise ratio in the spectrum, sacrificing spatial resolution in the process but improving the clarity of their data.
They then used this clearer dataset to test a machine learning technique called “non-negative matrix factorization” (NMF). NMF is a mathematical strategy that helps find patterns in data by dividing a huge dataset into smaller, more manageable pieces and making sure each portion is non-negative. As a result of their strategy, they had large pixel counts and fuzzy photos, but they also obtained good spectral data.
They then performed the NMF method on the original high-resolution dataset to retain detailed geographical information while initializing the factorization using the previously found spectral components. Finally, they integrated the findings of both methods to create a high-quality dataset with great spectral fidelity and spatial resolution.
The researchers verified PSNMF using synthetic data produced using a lab-created modeling method. These data simulated real-world difficulties, such as examining mineral samples created in harsh environments. The technique demonstrated remarkable efficacy in precisely recognizing and segregating various components, including those in minute quantities.
PSNMF was used to effectively separate and quantify overlapping components in real samples, such as a nanomineral and a nanocatalyst. This precise analysis requires understanding and creating new technologies that depend on these intricate nanostructures.
PSNMF represents a noteworthy advancement in chemical analysis at the nanoscale. This technique improves our understanding of nanomaterials and their applications in various disciplines, including sophisticated electronics and medical devices, by producing reliable results even in noisy data and overlapping signals.
EPFL funded the study.
Journal Reference:
Chen, H., et. al. (2024) Leveraging Machine Learning for Advanced Nanoscale X-ray Analysis: Unmixing Multicomponent Signals and Enhancing Chemical Quantification. Nano Letters. doi.org/10.1021/acs.nanolett.4c02446