R
EN 中
Back to projects
•
1 min read
GC-MS Signal Deconvolution & Identification Engine

87 monomers · 74,820 samples · ~99% blind-test accuracy (synthetic)

  • Python
  • 1D-CNN
  • PyTorch
  • GC-MS
  • Synthetic Data

Enterprise-collaboration exploration: an end-to-end 1D-CNN that deconvolves overlapping GC-MS signals of complex essential-oil mixtures and predicts the blend ratio of core raw materials — replacing manual expert peak-matching.

Approach & outcome

  • Synthetic-data baseline: scaled from 10 → 87 monomers, all pairwise combinations → 74,820 samples.
  • Reached ~99% identification accuracy on synthetic blind tests.
  • Archived due to the lack of real enterprise data and a Sim2Real domain gap — a first-hand lesson that data quality is the crux of landing industrial AI.