ADVENT
The olfactory data layer  ·  GraphAISense by ADVENT

AI learned to read, see and hear from data that already existed. Smell has none.

There is no archive of odour to train on. It has to be measured — on site, in the air that actually matters. ADVENT builds the instrument that makes smell measurable and the models that read it, so olfaction becomes a data layer instead of a missing sense.

GraphAISense Mk.2E graphene AI electronic nose, front three-quarter view
Olfaction is the sense machines still lack
30 – 38
Active sensing channels
≤ 2 s
Full-channel measurement cycle
< 1 s
On-device inference
1 ppm
Limit of detection (0.1 ppm on selected targets)

What is an electronic nose

An electronic nose recognises a smell the way your nose does — by its pattern.

Your nose does not measure one chemical. Around 400 types of olfactory receptors respond to a smell together, and the brain recognises the pattern. An AI electronic nose works the same way: an array of differently treated sensing cells responds to the air at once, and an AI model recognises that pattern as a specific odour — smouldering cable, a wastewater leak, a batch that has drifted. A single gas sensor, by contrast, reports one number for one gas and cannot tell a mixture apart.

Diagram comparing the human nose, a single gas sensor and an AI electronic nose

The instrument

GraphAISense Mk.2E

The shipping instrument. A graphene multisensor chip reads the mixture across 30 to 38 channels at once, and the Model Pack that interprets it runs on the instrument itself — measurement, recognition and local alarm complete at the edge.

GraphAISense Mk.2E graphene AI electronic nose

Multisensor chip

EN-MSC.3E graphene multisensor chip

Sensing channels

30 to 38, depending on configuration

Measurement cycle

2 seconds or less, all channels

AI inference

Under 1 second, on the instrument — no GPU, no cloud

Limit of detection

1 ppm general; 0.1 ppm for selected analytes under verified conditions

Air sampling

Built-in pump · 0.2 to 1.5 L/min · automatic solenoid flow control

Operating conditions

−5 to +50 ℃ · 20 to 98 %RH, non-condensing

Dimensions · power

270 × 250 × 90 mm · AC 220 V · 75 W or less

Connectivity

Ethernet · Wi-Fi · LTE · USB / MQTT over TLS · REST · OPC UA · Modbus TCP/RTU

Alert states

NORMAL · WARNING · ALARM, confirmed by continuous judgement

Mk.2E as published in the 2026 catalogue. Performance figures are ADVENT’s own verification results under the stated conditions, not accredited third-party certificates.

See the full Mk.2E specification
The platform

How an odour becomes a dataset.

Text, image and audio models arrived with archives already behind them. Odour had none. Every usable olfactory dataset has to be produced — measured under known conditions, labelled against a real event, and kept comparable across sites and across time.

Diagram: Measure, Fingerprint, Model Pack and on-device decision, with field data feeding back into the next Model Pack
Illustration: gas molecules landing on the sensing cells of a graphene chip
Step 01

Measure

A graphene multisensor chip reads the mixture across 30 to 38 channels at once, on a repeatable cycle. One reading is a vector, not a number — which is what makes it trainable at all.

Illustration: an odour mixture becoming a multi-channel fingerprint
Step 02

Fingerprint

Those vectors become an odour fingerprint — a pattern that stays comparable between batches, between sites and across seasons, with temperature and humidity accounted for. Comparable is the whole point: data that drifts is not data.

Illustration: one GraphAISense instrument connected to many Model Packs
Step 03

Model Pack

Fingerprints train a Model Pack for the question at hand, and it runs on the instrument itself. No GPU, no cloud, and the raw data does not have to leave the site to be useful.

The odour fingerprint

One measurement is not a number. It is thirty-eight of them.

A single gas sensor returns a scalar — did some value cross a threshold. Because every sensing cell on the graphene chip is functionalised differently, one exposure produces a different response on each channel at the same moment. That distribution is the fingerprint, and it is the vector, not the scalar, that makes a mixture separable.

Per-channel response, substance AResponse contour, substance B

Schematic, drawn to explain the principle. Not measured data.

CH 01SENSING CHANNELSCH 38

The recognition algorithm

From a 38-channel response to a decision.

Each exposure is read two ways: as a shape across channels (polar view) and as a response over time (2D view). The resulting vector is corrected, reduced with PCA, separated by class with LDA and classified with SVM — then weighed against persistence to give one of three alert states: NORMAL, WARNING or ALARM.

Polar chart of a 38-channel odour fingerprint comparing two odoursLine chart of channel responses through baseline, exposure and recovery
Recognition pipeline: raw signal, pre-processing, PCA, LDA, SVM and alert state

Schematic, drawn to explain the principle. Not measured data.

Why this data is different

Odour data cannot be scraped. It can only be collected.

The web wrote the corpus that language models learned from, and photographed the one that vision models learned from. Nothing on the internet smells. Every olfactory dataset that exists was produced by an instrument someone placed in a real location — so this data accrues only to whoever is measuring. It cannot be crawled, bought in bulk, or convincingly synthesised.

Illustration: the internet holds text, images and sound but no smell; odour data is collected on site

There is no corpus to start from

No archive to download, no public benchmark large enough to carry a product. Everyone starts at zero. What separates one position from another is how many hours of real, labelled air have been collected — and whether those hours are comparable to each other.

It compounds with every installation

An installed instrument keeps measuring the same place through seasons, shifts and process changes. That record is what turns a first deployment into a Model Pack, and a Model Pack into one that holds up on the next site. Data collected today makes the next deployment cheaper.

It stays where it is made

Inference runs on the instrument, so the raw signal does not have to leave the site to be useful. The operator keeps their data. That is what makes restricted places — plants, care facilities, public infrastructure — reachable at all.

This is why ADVENT builds the sensing material, the chip, the instrument and the model together. They are not four products. They are four stages of one data pipeline, and a company that outsources any of them cannot change the chemistry and the model at the same time.
Model Packs

One instrument. Hundreds of Model Packs.

A Model Pack is software, not a variant of the machine. Entering a new question — a new material, a new failure mode, a new site — means collecting data and training a pack, not redeveloping the instrument. Five areas are in the field today; the architecture puts no ceiling on the number, and the hardware cost base does not grow with it.

Diagram: the same GraphAISense hardware loads Model Packs from a library for safety, environment, industrial quality, healthcare research and robotics
Validation

Evidence stays scoped to the test that produced it.

These are ADVENT’s own verification results under the stated conditions. They are not accredited third-party test certificates, and neither result should be generalized into an overall accuracy figure.

January 2026

Smouldering wire-insulation test

Smouldering wire insulation as the target, against chlorine, ethanol and isopropyl alcohol as three background chemicals. The target was correctly identified across that test set.

Q4 2025

Blind odour-class test

Chemical, wastewater and sewage odour classes separated with high accuracy, and a mixed sample resolved into its two components.

Field trials with public-sector partners

Korea Environment Corporation (K-eco)

Diagram of the October 2025 odour source-tracking field test

October 2025

Odour source tracking for residential complaints

Trained on four source odours (chemical/paint, textile, night-soil and sewage treatment) and two mixtures; three unlabelled samples were traced back to their sources.

Incheon Port Authority

Incheon Port Authority AI e-nose field trial with a security patrol robot

2026 · In progress

AI e-nose on a security patrol robot and ESS

Port passenger terminal: an e-nose built into a SEO Robotics patrol robot, alongside fixed units at energy-storage and solar facilities.

Gyeonggi-do

Gyeonggi-do on-device AI and graphene e-nose fire detection field trial

2026 · In progress

On-device AI fire detection with graphene e-noses

Gyeonggi-do with Suwon, Hwaseong and Icheon: e-noses linked with AI CCTV and edge AI to catch fire precursors in industrial complexes and markets.

Partners are named to identify where each field trial runs. Naming does not imply endorsement, and these are not purchase or delivery records.

Read the full validation and deployment record