SmellNet

SmellNet is a comparatively large dataset for sensor-based machine olfaction: real-world smell measurements collected from a compact array of low-cost metal-oxide (MOX) gas sensors.

The dataset supports:

• single-substance recognition from multichannel sensor time series
• mixture distribution prediction from odor mixtures
• cross-modal learning with paired GC-MS-derived chemistry priors
• research on temporal modeling for smell sensing

The accompanying paper introduces ScentFormer, a Transformer-based model for smell recognition and mixture prediction.

Highlights

• ~828,000 time-series readings
• 50 base substances across nuts, spices, herbs, fruits, and vegetables
• 43 mixtures with fixed ingredient volumetric ratios
• ~68 hours of controlled acquisition
• headline results from the paper:
  • 63.3% Top-1 accuracy on SmellNet-Base with GC-MS supervision
  • 50.2% Top-1@0.1 on seen mixtures

Repository structure

This Hugging Face repo contains the dataset assets used in the SmellNet project.

• base_data/
  Curated base-substance sensor data used for the main paper experiments.
  This corresponds to the SmellNet-Base task.

• mixture_data/
  Mixture sensor data for predicting normalized ingredient ratios over 12 odorants.
  This corresponds to the SmellNet-Mixture task.

• gcms_processed/
  Processed GC-MS-derived features and chemistry priors used for cross-modal learning.

• gcms_data/
  Raw or intermediate GC-MS-related files used to construct chemistry-aware representations.

Tasks

1. SmellNet-Base

The base task focuses on single-substance classification from multichannel sensor time series.

• 50 substances
• 5 broad categories:
  • nuts
  • spices
  • herbs
  • fruits
  • vegetables
• controlled repeated sensing sessions across different days

2. SmellNet-Mixture

The mixture task focuses on mixture composition prediction from sensor time series.

• 12 odorants
• 43 mixture configurations
• prediction target is a normalized ingredient-ratio vector

Sensor setup

The SmellNet paper uses low-cost portable MOX sensors.

For the base-substance setting, the reported experiments use 6 sensor channels.

For the mixture setting, the reported experiments use 4 Grove channels collected at a higher sampling rate.

These signals are best interpreted as relative sensor responses and temporal dynamics, rather than calibrated absolute chemical concentrations.

What is included here

This dataset repo is meant to host the data assets.

For:

• training code
• preprocessing code
• modeling code
• analysis scripts
• paper figures

please see the companion GitHub repository:

GitHub: https://github.com/MIT-MI/SmellNet

Intended use

SmellNet is intended for research on:

• machine olfaction
• temporal representation learning
• multimodal learning with chemistry priors
• mixture prediction
• low-cost sensor-based perception

Potential applications include food analysis, environmental monitoring, healthcare-related sensing research, and interactive multimodal systems.

Notes and limitations

• Data were collected in a controlled environment, so real-world generalization to unseen environments may be harder.
• Sensor channels are manufacturer-labeled responses and are not direct measurements of pure analytes in food headspace.
• GC-MS information is used as an ingredient-level chemistry prior, not as a per-sample test-time input.
• Performance on unseen mixtures remains significantly harder than on seen mixtures.

Citation

If you use this dataset, please cite:

@inproceedings{feng2026smellnet,
  title={SmellNet: A Large-Scale Dataset for Real-World Smell Recognition},
  author={Feng, Dewei and Dai, Wei and Li, Carol and Pernigo, Alistair and Liang, Paul Pu},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}