Elly Sufriadi, Vattany Al-Asra Imran, Lelifajri Lelifajri, Amrin Barata, Hilmi Sulthanah, Irvanizam Irvanizam, Rinaldi Idroes
Adulteration of patchouli oil with cheaper substitutes such as gurjun balsam oil (GBO), castor oil (CO), and palm oil (PMO) is a persistent problem in the essential-oil trade. This study presents an interpretable machine-learning framework for detecting patchouli oil adulteration using attenuated total reflectance Fourier-transform infrared (ATR-FT-IR) spectroscopy. A total of 375 spectra were acquired from certified reference material, authentic patchouli oil, pure adulterants, and binary mixtures spanning 20 concentration levels from 0.5 to 10.0% v/v. Six preprocessing strategies (Raw, SNV, MSC, Savitzky–Golay first and second derivatives, min–max normalization) were combined with two spectral domains (full spectrum 4000–499 cm−1 and fingerprint region 1800–600 cm−1), and four classifiers (Logistic Regression, k-nearest neighbours, Random Forest, Gradient Boosting) were evaluated under group-aware nested cross-validation in which all replicate spectra of a physical sample were kept within the same fold to prevent replicate-driven leakage. Under this leakage-aware design, binary classification was near-perfect but not uniformly perfect: palm oil reached F1 = 1.000 across all classifiers, gurjun balsam oil ranged from 0.967 (Logistic Regression) to 1.000 (k-nearest neighbours), and castor oil plateaued at F1 = 0.978 across all classifiers, with bootstrap 95% confidence intervals of [1.000, 1.000], [0.984, 1.000], and [0.951, 0.995], respectively. Group-aware permutation testing (null F1 ≈ 0.44; p ≈ 0.005), learning curves, and Monte-Carlo cross-validation supported the legitimacy of these associations within the present controlled dataset, while cross-preprocessing and cross-adulterant analyses (mean transfer F1 ≈ 0.44) revealed adulterant-specific signatures rather than universal transferability. A strict leave-low-concentration-out test confirmed detection of low-level adulteration (recall ≥90% at 0.5–2.0% v/v). SHAP-based interpretation localized the discriminative information in chemically meaningful regions (carbonyl, C-H, and skeletal vibrations); a region-exclusion analysis showed that the GBO model retains high performance after removing the atmospheric CO₂ region (F1 from 1.000 to 0.976), indicating only a minor, recoverable reliance on that region. The framework offers an interpretable, internally validated approach under controlled binary-mixture conditions and a robust analytical baseline for future inter-laboratory and real-market validation. © 2026 Elsevier B.V.
Department of Chemistry, Faculty of Mathematics and Natural Sciences, Universitas Syiah Kuala, Banda Aceh, 23111, Indonesia; BPS-Statistics Indonesia, Southeast Sulawesi, Kendari, Indonesia; Department of Statistics, Faculty of Mathematics and Natural Sciences, Universitas Syiah Kuala, Banda Aceh, 23111, Indonesia; Department of Informatics, Faculty of Mathematics and Natural Sciences, Universitas Syiah Kuala, Banda Aceh, 23111, Indonesia; Department of Pharmacy, Faculty of Mathematics and Natural Sciences, Universitas Syiah Kuala, Banda Aceh, 23111, Indonesia; Atsiri Research Center (ARC), PUI-PT Nilam Aceh, Universitas Syiah Kuala, Banda Aceh, 23111, Indonesia
Research at a Glance
Register to unlock
Topics & SDG Alignment
Register to unlock
Collaboration
Register to unlock
Author Profile (Selected)
Register to unlock
References Overview
Register to unlock
Journal & Source
Register to unlock
Metadata & Integrity
Register to unlock