How machines detect chemical patterns in air.
Every environment carries an invisible chemical signature. Machine smell is the study of reading that signature — turning the reaction of a sensor array into a stable, recognizable pattern.
What we research
We do not try to identify every molecule one by one. We read the combined response of a low-cost sensor array as a fingerprint — a pattern that a model can learn, the way a person recognizes coffee without naming each compound in it.
Technical terms: volatile organic compounds (VOCs), gas sensing, electronic nose, sensor array.
The array we are building
Our electronic-nose design combines commercially mature, low-power sensor families, chosen so no single sensor is a crutch.
The full Phase 1 rig runs multiple units of each family behind an I2C multiplexer, with diaphragm pumps and PTFE tubing for controlled sampling.
The fingerprint
The array's raw output is a tangle of curves. A fingerprint is what survives processing: a stable, comparable pattern a model can learn and a protocol can reproduce. Getting there is signal processing, sensor fusion, baseline normalization, and feature design — the step between raw electronics and machine-readable smell.
The fingerprint is the unit everything downstream depends on, so it has to be stable across runs before any claim rests on it. The feature shapes and data contracts it relies on are already fixed; what the bench is establishing now is whether real hardware produces fingerprints that hold up run to run.
Technical terms: signal processing, feature extraction, sensor fusion, baseline normalization, time-series representation.
Where this stands
The bench is live: the ESP32-S3 is flashed and three sensor families — BME688, SGP40, SHT40 — are returning real readings, with MiCS-6814 in bring-up. We are completing the rig and characterizing how the array responds run to run, before any detection claim rests on it.
The open question — the reason the lab exists — is whether clean-air baselines and controlled aerosol events separate repeatably on real hardware, across humidity and background air. The bench records controlled exposure runs toward exactly that Bench — baselines, blanks, and confounder controls alongside every exposure. These are recorded runs, not results: the analysis that would answer the question is not yet gated.
Further reading
- →Controlled sniffing — how air reaches the array in a repeatable way.
- →On-device AI — turning the fingerprint into a decision on the board itself.
- →Truth gates — how each result is staged and what it is allowed to claim.