Leveraging Generative AI for Denoising Instrumental Analysis Data

Signal- and noise-related concepts originated in communications a century ago and entered analytical chemistry in the 1960s with the computer revolution. The long-standing goal of reducing the limit of detection, or figuratively “hearing the grass grow,” remains unmet. Noise is inherently intrinsic to every chemical signal and can ultimately hinder analysis at low concentrations or when the analyte signals are weak. Consequently, most chemical measurements still rely on denoising during acquisition with digital filters. Many such methods were developed between 50 and 70 years ago, such as the popular Savitzky-Golay method and moving averages. They perform well in conventional cases; however, their limitations, including peak broadening and weak suppression of low-frequency noise (such as baseline wander, pink noise, and periodic interference), are well-documented. In parallel, imaging sciences have made substantial progress in restoring blurry or noisy data. Can we apply those ideas to 1D signals? Generative AI provides a route to translating these developments into one-dimensional analytical signals, including chromatography, capillary electrophoresis, mass spectrometry, microwave spectroscopy, optical spectroscopy, and nuclear magnetic resonance, with great success. This talk focuses on how analytical chemists can benefit from generative AI and implement advanced denoising approaches on their own data, such as denoising by optimization formulations, the Perona-Malik method, bilateral, and guided filters for challenging noisy cases encountered in analytical chemistry. We will also discuss practical tips for smoothing instrumental output. Pertinent case studies will be discussed in relation to relevant analytical data.