External reflectance Fourier-transform infrared (ER-FTIR) spectroscopy is a key non-destructive archaeological analysis tool. However, the resulting spectra are often distorted by surface scattering and complex reflectance backgrounds. This study presents an automated machine learning pipeline designed to classify mineralogical and organic samples with high diagnostic precision. Leveraging a dataset of 189 diverse samples, a robust Random Forest (RF) classifier integrated with advanced chemometric preprocessing is effectively implemented. Specifically, the application of a first-order Savitzky–Golay derivative and regional spectral slicing (1800–400 cm⁻¹) is evaluated against raw intensity models. The optimized pipeline achieves a significant overall accuracy matching human expert performance benchmarks. Key findings demonstrate substantial diagnostic power for organic binders or coatings and silicate groups. Furthermore, it successfully isolates characteristic vibrational inflections such as Si-O-Si stretching and Carbonyl overtones from complex backgrounds. However, heterogeneous unclassified category materials present challenges due to high chemical variance. The model proves to be reliable for the identification of primary archaeological materials. This methodology enables the field identification of diverse materials, such as silicate-based pigments (e.g., distinguishing between lapis lazuli, kaolinite, and orthoclase via Si-O-Si stretching), carbonates like calcite or marble in sculpture, and complex polymineralic rocks such as granite or breccia. This technique enhances characterization clarity by isolating subtle diagnostic chemical footprints that are typically obscured by surface scattering and complex background distortion. The integration of derivative-based preprocessing with ensemble learning provides a scalable and objective framework for real-time field identification. This AI-driven approach enhances the speed and consistency of cultural heritage conservation and material characterization by reducing the over-reliance on manual expert interpretation in the field.