Skip to content

Workflow: Domain Templates (Meat, Microbial, and Beyond)

📋 Standard Header

Purpose: Provide pre-configured workflow templates for specific food domains (meat, dairy, grains, microbial ID) with sensible defaults.

When to Use: - Quick-start authentication for common food matrices (meat species, microbial strains) - Leverage domain-specific preprocessing and feature defaults - Compare results to published domain benchmarks - Adapt existing validated workflows to similar matrices - Training users on standardized approaches for specific domains

Inputs: - Format: Same as oil authentication (HDF5/CSV with spectra) - Required metadata: Domain-specific labels (e.g., meat_type, microbial_species, grain_variety) - Optional metadata: source, treatment, preparation_method - Wavenumber range: Domain-dependent (meat: 600–1800 cm⁻¹; microbial: 800–1800 cm⁻¹) - Min samples: 50+ per class (same as general authentication)

Outputs: - Same as oil authentication workflow (confusion matrix, PCA, metrics, report) - Domain-specific interpretation notes in report

Assumptions: - Domain template validated on representative samples from target domain - Sample preparation matches template specifications (powder, liquid, cuvette type) - Spectroscopy method matches template (Raman vs FTIR) - Within-domain variability captured in training data (varieties, sources, seasons)

What this chapter covers

  • How domain templates map to the oil-auth style pipeline (preprocessing + classifier).
  • Typical metadata/label expectations per domain (meat_type, species/strain, etc.).
  • When to use a domain template vs configure your own workflow.
  • Links to meat/microbial tutorial pages for runnable examples.

Outline

  • Template concept: Thin wrappers around preprocessing + classification; default features/models.
  • Meat: Raman/FTIR use cases; label expectations; adapting oil defaults.
  • Microbial: Spectral IDs; class imbalance considerations; QC steps.
  • Dairy/adulteration (future): Apply the same preprocessing/ratios/PCA + classifier pattern; record instrument (FTIR/NIR), matrix (milk powders/liquids), target labels (adulterant level/type); reuse reproducibility fields for plots/reports.
  • Spices/grains (future): Heterogeneous matrices; emphasize preprocessing choices (baseline, normalization), feature selection (key bands), and QC/statistics similar to oil workflows.
  • Extensibility: Adding new domain templates; using CLI domains command (if applicable).
  • Pointers: See ../meat_tutorial.md and ../microbial_tutorial.md for code/CLI recipes.

When Results Cannot Be Trusted

⚠️ Red flags for domain-specific templates:

  1. Domain template applied without verifying it covers sample diversity
  2. Template trained on limited subset of domain; real samples more variable
  3. Boundary cases (organic vs conventional, rare varieties) not represented

  4. Fix: Include diverse sources, varieties, and processing methods in template validation

  5. Spectroscopy method mismatch (template for Raman applied to FTIR data)

  6. Different methods give different spectra; models don't transfer without retraining
  7. Spectral ranges, baseline, and peak positions different

  8. Fix: Use method-specific template; validate transfer before cross-method deployment

  9. Sample preparation not matching template assumptions (template assumes dried powder, new samples are liquid)

  10. Preparation dramatically affects spectra; cuvette, path length, temperature critical
  11. Template model won't work if prep fundamentally different

  12. Fix: Match sample prep to template specifications; retrain if prep changes

  13. Seasonal or temporal variation not addressed (template trained in summer, deployed in winter)

  14. Ambient temperature, storage time, ripeness, and harvest effects not captured
  15. Spectra may shift seasonally, violating template assumptions

  16. Fix: Include samples from different seasons/harvest times; validate temporal generalization

  17. Reference database for domain template incomplete (missing adulterant types, new varieties)

  18. Template can only detect adulterants in training set
  19. Novel adulterant or variety will be misclassified

  20. Fix: Continuously update reference database; validate on new adulterants before deployment

  21. Domain template assumes homogeneous matrix (oil domain applied to olive oils, but different cultivars have very different compositions)

  22. Intra-domain variability can exceed inter-domain differences
  23. Single template insufficient

  24. Fix: Stratify by important factors (variety, origin, processing); use multiple templates or include factors in model

  25. No cross-validation across domain subsets (all training on one supplier/region)

  26. Supplier-specific patterns learned; won't generalize
  27. Cross-source/region validation reveals true robustness

  28. Fix: Include multiple suppliers/regions in training; validate on held-out sources

  29. Domain boundaries unclear (when is sample "in domain" vs "out of domain"?)

  30. No objective rule for when template applies; user confusion
  31. Can lead to inappropriate use
  32. Fix: Define domain explicitly (e.g., "olive oils from Mediterranean, post-harvest, stored <2 years"); flag out-of-domain samples

Next steps

  • Use a template for rapid prototyping; switch to custom pipelines for specialized datasets.
  • Explore Protocols & reproducibility to document template use in studies.