*Result*: AI-assisted methodology for robust digital measurements by Raman spectroscopy: Quantification of inorganic pollutants in water.

*Raman spectroscopy is a versatile analytical tool, yet it often struggles with low sensitivity, hardware noise, and environmental interference. To address these limitations, this study presents an automated, Artificial Intelligence (AI)-assisted methodology to convert noisy optical signals into robust digital measurements. The process involves acquiring high-dimensional, noisy spectral data from analyte solutions. A grid search across various algorithms identifies the optimal pre-processing pipeline to minimize noise variance and ensure metrological repeatability. Instead of relying on raw sensor feeds, the method fits a Gaussian curve combined with a polynomial baseline to the data, extracting precise measurements from the peak of this mathematical model. Supported by AI, the method successfully separates multiple optical signals and their shifts originating from interactions among analytes, proving itself capable to compensate also for possible hardware misalignment and thermal drift. As such, it can be used to quantify the concentration of selected inorganic pollutants in a mixture of analytes. The primary application addressed in this work is quantifying inorganic pollutants in water, to enable in situ analysis without continuous expert supervision. Tests on binary and ternary mixtures of inorganic pollutants in pure water demonstrated that the Mean Absolute Percentage Error (MAPE) for nitrate was consistently below 10% in the concentration range between 0 mg/L to more than 15 000 mg/L, dropping to under 5% for concentrations exceeding 1000 mg/L. For concentrations below 1000 mg/L, the Mean Absolute Error (MAE) values were 67 mg/L for nitrate, 1475 mg/L for sulfate, and 736 mg/L for nitrite, respectively. • Quantification of concentration of pollutants by traditional Raman spectroscopy. • Processing of spectra to minimize sample standard deviation of detected intensity. • Feature extraction protocol from the main peak of each pollutant. • Use of feature-based and interpretable machine learning. • Mean Absolute Error of less than 5% for nitrate concentration above 1000 mg/L. [ABSTRACT FROM AUTHOR]*