The challenge is real and surprisingly difficult to detect. While green coffee beans can often be told apart by eye, roasting and grinding erase the morphological features that distinguish the two species. Once beans are roasted, authentication requires chemistry. Traditional approaches rely on trained cuppers whose sensory evaluations are subjective and generally fail to detect Robusta adulteration until it exceeds roughly 35 percent of the blend. Laboratory methods such as high-performance liquid chromatography offer objectivity but demand expensive instruments, skilled operators, and long analysis times, with per-sample costs around 100 US dollars. Nuclear magnetic resonance spectroscopy can quantify marker compounds like 16-O-methylcafestol, a Robusta-specific molecule, but remains too slow and costly for routine screening.

The Taiwanese team, led by Bing-Chen Gu and Chia-Che Wu, took a different route. Their platform uses disposable screen-printed sensor chips with a three-electrode configuration: carbon electrodes to measure caffeine, chlorogenic acids, and trigonelline; a gold electrode for sugar-related measurements; and a silver/silver chloride reference electrode. After a small aliquot of brewed coffee is mixed with species-specific electrolytes, the device applies voltammetric techniques that exploit well-characterized reaction pathways. Caffeine and chlorogenic acids are oxidized at the electrode surface in the low-potential region around 200 to 300 millivolts. Trigonelline is first converted under alkaline conditions to electroactive nicotinic acid, which then produces a prominent oxidation peak near 1500 millivolts. Reducing sugars are detected through copper redox chemistry, in which the sugars reduce copper ions to cuprous oxide under alkaline conditions, generating a characteristic signal in the negative potential region near minus 250 millivolts.

The entire electrochemical measurement takes less than three minutes per sample once the coffee extract is prepared, using only microliter-scale volumes. The team validated the platform rigorously, establishing linear calibration over working ranges of 20 to 500 milligrams per deciliter for caffeine and chlorogenic acids, 10 to 100 for trigonelline, and 65 to 1250 for reducing sugars. Limits of detection ranged from 1.26 milligrams per deciliter for trigonelline to 19.27 for reducing sugars. Spike-recovery experiments in real coffee matrices returned recoveries between 97.2 and 110.3 percent for trigonelline and 92.9 to 113.9 percent for sugars, comfortably within the accepted 80 to 120 percent window. Testing across 15 independently manufactured sensor chips from three production batches yielded relative standard deviations between 6.23 and 9.17 percent, demonstrating that the disposable chips perform consistently despite batch-to-batch variation. Matrix effects, the distortion of signals by other compounds in the coffee, were minimal, with slope ratios remaining close to 100 percent.

Crucially, the electrochemical readings captured the expected chemistry of coffee roasting. Across representative samples, chlorogenic acids, trigonelline, and reducing sugars all declined progressively as roast level deepened, reflecting thermal degradation, Maillard reactions, and caramelization. Caffeine, by contrast, showed only weak dependence on roast level, and Robusta consistently maintained higher caffeine concentrations than Arabica, matching decades of chromatographic evidence. These roast-dependent trends confirmed that the portable sensor was generating chemically meaningful data rather than artifacts, providing a credible foundation for the machine learning stage.

For the classification task, the researchers assembled a dataset of 138 samples: 66 Arabica coffees spanning 11 varieties from Ethiopia, Kenya, and Colombia; 18 Robusta samples from Vietnam, India, and Brazil; and 54 experimentally prepared blends at ratios ranging from 95:5 to 30:70 Arabica to Robusta. All beans were roasted to controlled light, medium, and dark levels verified by Agtron measurements, and blends were formulated from beans roasted to matching levels to prevent roast differences from confounding the analysis. Five variables, caffeine, chlorogenic acids, trigonelline, reducing sugars, and roast level, were fed into two machine learning classifiers. The random forest model achieved 97.37 percent test accuracy, while XGBoost classified every test sample correctly, achieving 100 percent accuracy with perfect precision, recall, and F1-scores for all three classes. Feature importance analysis revealed caffeine as the single most influential variable, followed by chlorogenic acids and roast level, though caffeine alone proved insufficient, misclassifying blends as pure Arabica when used without the other descriptors.

The team also stress-tested their models with stricter validation designs. In a leave-one-origin-out analysis, where all samples from one coffee origin were withheld from training, the binary Arabica-versus-Robusta discrimination retained perfect accuracy, balanced accuracy, and macro-F1 of 1.000, suggesting the models generalize to origins never seen during training. External validation using compositional data extracted from previously published studies, comprising 17 Arabica and 3 Robusta samples, also produced 100 percent correct classification by the trained XGBoost model.

Predicting blend ratios quantitatively proved harder than simple classification, and the researchers were candid about the pitfalls. A neural network regression model trained on the same five features achieved a coefficient of determination of 0.980 on a randomly split test set, with root mean square error of 0.050 and mean absolute error of 0.032. But because random splitting can scatter samples from the same Arabica-Robusta pairing across both training and test sets, the team implemented a far more demanding leave-one-blend-pair-out validation, in which all 18 samples from one complete blend pairing were withheld from every stage of model development. Under this nested group-aware framework, which selected an ensemble of four neural network architectures from a pool of 48 candidates per fold, pooled performance remained strong: R-squared of 0.9121, root mean square error of 0.0676, and 47 of 54 held-out predictions, or 87.04 percent, falling within plus or minus 0.10 of the true blend fraction. Pair-specific R-squared values ranged from 0.8863 to 0.9353, with the largest errors concentrated at low blend ratios for one particular pairing. The drop from 0.980 to 0.9121 confirmed that random splitting had provided an optimistic estimate, while the retained performance demonstrated genuine transferability to unseen blend combinations.

The interpretability analysis added further confidence. SHAP analysis indicated that caffeine and trigonelline were the most influential variables driving the neural network’s blend-ratio predictions, with chlorogenic acids, sugars, and roast level contributing smaller but systematic effects. The model, in other words, relied on the full multivariate chemical fingerprint rather than gaming any single feature. Benchmark comparisons underscored the value of the neural network approach: multiple linear regression managed only an R-squared of 0.146, showing that simple linear combinations cannot capture the nonlinear relationships among composition, roasting, and blend ratio, while random forest regression reached 0.951 but remained limited by its piecewise, tree-based structure.

The implications extend well beyond coffee fraud. The platform requires only a small extract aliquot, completes measurement within three minutes, and operates on portable, potentially low-cost hardware, making it suitable for real-time quality control at roasteries, import checkpoints, and retail settings. The authors acknowledge limitations: the blended samples used components with similar roast levels, so the current model is validated primarily for matched-roast blends, and broader sampling across origins, processing methods, and harvest seasons will be needed before wider deployment. Independent external validation of blend-ratio regression also remains outstanding because no published dataset contained all five required variables. Still, the study demonstrates a compelling template for intelligent food authentication, one that could readily extend to other beverages and foodstuffs by expanding the target analytes and sample diversity. As adulteration schemes grow more sophisticated, the marriage of cheap disposable sensors and machine learning may prove one of the most practical defenses the food industry has.

Subject of Research: Portable electrochemical sensing with machine learning for coffee species authentication and blend ratio prediction

Article Title: Portable electrochemical sensing coupled with machine learning for rapid Arabica and Robusta coffee authentication and blend ratio prediction

Article References: Gu, B.-C., Chung, K.-J., Hong, Z.-W., Yang, N.-H., & Wu, C.-C. (2026). Portable electrochemical sensing coupled with machine learning for rapid Arabica and Robusta coffee authentication and blend ratio prediction. Food Chemistry: X, 39, Article 104523. https://doi.org/10.1016/j.fochx.2026.104523

Image Credits: AI Generated

DOI: 10.1016/j.fochx.2026.104523

Keywords: coffee authentication, Arabica, Robusta, electrochemical sensing, machine learning, food adulteration, XGBoost, artificial neural network, caffeine, chlorogenic acid, blend ratio prediction, portable biosensor