Predictive Classification of Contaminants Encountered During the Distillation of Shochu,a Distilled Beverage Native to Japan

Significance of the Topic

Shochu is a traditional Japanese distilled spirit with significant economic and cultural value. Ensuring its purity and quality during production is critical to maintain consumer trust and meet regulatory requirements. Detecting trace contaminants introduced during distillation supports robust quality control and prevents costly recalls.

Objectives and Study Overview

This study aimed to develop a predictive sample class model (SCP) to classify and detect common contaminants encountered in Shochu distillation. Contaminants evaluated included detergents, rubber glove residues, and machine oil. The investigators prepared both uncontaminated and spiked Shochu samples from Osaka, Tokyo and San Jose sources, then evaluated multivariate data analysis approaches for discrimination and prediction.

Methodology and Instrumentation

Sample Preparation and Extraction:

• 1.0 g Shochu spiked with defined amounts of detergent, rubber glove or machine oil.
• Solid-phase microextraction (SPME) at 40 °C for 40 min (100 µm PDMS fiber).

GC-MS Analysis:

• Gerstel MPS2 autosampler with Agilent 7890 GC/5975C MSD system.
• Column: 10 m × 0.18 mm × 0.18 µm DB-Wax.
• Splitless injection, oven program 35 °C (2 min) → 240 °C at 30 °C/min.
• Scan mode 35–450 u, EI 70 eV, ion source 230 °C, transfer line 240 °C.

Data Processing:

• AMDIS for deconvolution and component extraction.
• Agilent Mass Profiler Professional for alignment and entity list generation (~2 376 features).
• Feature filtering in three steps: frequency (100 % in at least one class), ANOVA p < 0.05, coefficient of variation < 45 %.

Multivariate Modeling:

• PCA for exploratory visualization.
• SCP algorithms: Decision Tree (DT), Naïve Bayes (NB), Neural Network (NN), Partial Least Squares Discriminant (PLSD) and Support Vector Machine (SVM).

Main Results and Discussion

PCA Score Plots under successive filters showed increasing separation of uncontaminated Shochu and different contamination classes. SCP performance on independent test samples revealed:

• Filter 3 (frequency + ANOVA + CV) yielded the strongest classification features.
• DT and SVM achieved 100 % accuracy under filter 3.
• PLSD maintained 90 % accuracy, demonstrating sensitivity to subtle chemical differences.
• NB and NN models showed lower robustness (50–73 % under strict filtering).

These results indicate a minimal subset of reproducible features drives accurate contamination prediction, and that appropriate filtering is critical for model success.

Benefits and Practical Applications of the Method

The developed SCP workflow offers:

• Rapid screening for trace contaminants in Shochu production, enhancing QA/QC throughput.
• Prevention of contaminated batch release and assurance of product consistency.
• A template for multivariate GC-MS data analysis applicable to other beverage or food matrices.

Future Trends and Utilization Possibilities

Advances may include:

• Real-time, in-line SPME-GC-MS monitoring integrated into distillation systems.
• Expansion to detect a broader range of process-related impurities and flavor markers.
• Automation of feature extraction and model retraining to adapt to seasonal or feedstock variations.
• Coupling with high-throughput data platforms and cloud-based predictive analytics.

Conclusion

This work demonstrates that targeted SPME-GC-MS combined with rigorous feature filtering and SCP algorithms can reliably classify Shochu samples by contamination type. Decision Tree and SVM models, supported by a concise set of reproducible entities, achieved perfect prediction accuracy under optimized conditions. The approach provides an effective QA/QC tool for the beverage industry.

Instrumentation

Gerstel MPS2 Multiple Purpose Sampler; Agilent 7890 GC/5975C MSD; J&W 121-7012LTMDB-Wax column; AMDIS 2.69; Agilent Mass Profiler Professional, ChemStation software.