SmellNet: A Large-scale Dataset for Real-world Smell Recognition

Overview of the smell sensing and modeling pipeline: hardware setup, multichannel sensor responses, and example model predictions.

Paper • Dataset • Poster • Code • Demo

SmellNet is a benchmark for sensor-based machine olfaction built from portable low-cost MOX gas sensors. It supports both single-substance recognition and mixture prediction, and introduces ScentFormer, a temporal model for learning from multichannel smell sensor time series with optional training-time GC-MS supervision.

Highlights

• 828K sensor timesteps
• 50 base substances across nuts, spices, herbs, fruits, and vegetables
• 43 controlled mixtures
• 68 hours of recordings
• Benchmark tasks for both classification and mixture-ratio prediction
• Optional sensor-chemistry alignment using ingredient-level GC-MS representations

Main Results

ScentFormer achieves:

• 63.3% Top-1 accuracy on SmellNet-Base with GC-MS supervision
• 50.2% Top-1@0.1 on seen mixtures
• 16.0% Top-1@0.1 on unseen mixtures

Main benchmark results: SmellNet-Base and mixture tasks.

These results show that temporal modeling is effective for smell recognition, while generalization to unseen mixtures remains a core challenge.

What is in this repository?

This repository contains:

• code for training and evaluating smell-recognition models
• preprocessing and analysis utilities for sensor data
• GC-MS processing pipelines for cross-modal supervision
• Arduino and data-collection utilities used in the sensing setup
• scripts for reproducing main experiments and regenerating figures

Large files (sensor recordings, gcms_processed/, etc.) are not stored in Git. Download them from the SmellNet dataset on Hugging Face and place them in the paths described in Quick Start so scripts can find them locally.

Benchmark Tasks

SmellNet-Base

A 50-way classification benchmark for recognizing single substances from multichannel sensor time series.

SmellNet-Mixture

A benchmark for predicting normalized mixture ratios over 12 base odorants, including:

• seen mixtures: known combinations under new recording sessions
• unseen mixtures: novel combinations for compositional generalization

Method Overview

The overall pipeline is:

1. collect multichannel sensor responses from portable MOX sensors
2. optionally apply temporal differencing to emphasize signal dynamics
3. segment the signal into sliding windows
4. train temporal models such as CNNs, LSTMs, and ScentFormer
5. optionally align sensor embeddings with ingredient-level GC-MS embeddings during training

Quick Start

1) Clone the repository

git clone https://github.com/MIT-MI/SmellNet.git
cd SmellNet

2) Install a minimal environment

pip install -r requirements.txt

This installs dependencies for training, analysis, and optional tools (serial capture, t-SNE/UMAP, Plotly). If you only need a minimal stack, you can install the core packages by hand (torch, pandas, numpy, scikit-learn, matplotlib). For a particular CUDA version of PyTorch, install torch from pytorch.org first, then run pip install -r requirements.txt (pip will keep your existing torch if it satisfies the constraint).

3) Download the dataset

Download the released files from Hugging Face — DeweiFeng/smell-net and unpack them into this clone. The dataset card lists archives and filenames.

Running Experiments

SmellNet-Base classification / contrastive training

cd models
python run.py \
  --train-dir ../base_data/training \
  --test-dir ../base_data/testing \
  --real-test-dir ../base_data/testing \
  --gcms-csv ../gcms_processed/gcms_food_vectors.csv \
  --models transformer \
  --epochs 90 \
  --batch-size 32 \
  --lr 0.001

For fuller sweeps, see:

• scripts/run_experiment.bash
• scripts/run_experiments.sh

Mixture prediction

Use:

• models/run_mixture.py

and the commented mixture template in:

• scripts/run_experiment.bash

Repository Structure

models/           training and evaluation code
scripts/          experiment wrappers and figure-generation utilities
analysis/         notebooks and analysis scripts
base_data/        SmellNet-Base sensor CSVs (training/testing; not in git — download from Hugging Face)
gcms_processed/   GC-MS vectors (not in git; download from Hugging Face)
gcms_analysis/    scripts to build GC-MS embeddings (outputs go to gcms_processed/)
data_collection/  acquisition helpers and sensor-side tooling
data_stats/       summary plots and tables
figures/paper/    regenerated paper figures
preprocessing/    raw-to-structured data utilities
Arduino/          firmware and sensor libraries

Applications

SmellNet can support research and prototyping in:

• machine olfaction
• food and beverage monitoring
• allergen-relevant sensing
• multimodal sensing
• environmental and industrial monitoring
• digital olfaction and human-AI interaction

Citation

If you use SmellNet in your work, 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}
}

Contact

For questions or collaboration, please contact Dewei Feng at dewei@mit.edu.