Industry · Gas detection

Gas detection & chemical sensing

One micro-hotplate, many gases, and a signal that quietly drifts every month it runs. Reading a stable concentration out of a metal-oxide sensor is an inverse problem in disguise, and it is the problem chemical sensing has to solve to be trusted in the field.

01The problem

Metal-oxide gas sensors cannot hold an absolute calibration. Sensirion, for instance, ships its SGP40 and SGP41 MOX sensors reporting a relative VOC Index rather than an absolute concentration, because the raw signal drifts and is continually re-referenced against the last 24 hours of ambient air. The physics is under-determined: many analytes fold into one slowly degrading signal, entangled with temperature and humidity.

Many analytes, one drifting signal. The signature problem of the field, using a commercial MOX sensor such as Sensirion's SGP40 as the example: several gases collapse onto a single resistance that slowly ages.

02How we solve it

Separating the analyte from the drift.

Agentriq treats calibration as a standing inference problem: model the sensor physics, learn the cross-sensitivities, and re-reference the signal continuously instead of trusting a factory number.

Temperature modulation makes one sensor an array. Pulsing the heater across its range produces gas-specific transient fingerprints, the virtual sensor array that replaces physical arrays.

Temperature modulation and virtual sensor arrays

Thermally pulsing the micro-hotplate across its operating range, roughly 225 to 350 C, turns one sensor into a time-resolved fingerprint, so a single element can discriminate gases an array once needed.

Drift compensation via domain adaptation

Domain-adaptation and knowledge-distillation models align today's feature distribution back to the trained one, holding classification accuracy as the baseline resistance ages.

Calibration transfer

Transfer maps port a trained model from a reference sensor to a fresh unit, avoiding a full recalibration of every device off the line.

Cross-sensitivity and environment correction

Explicit humidity and temperature terms strip the environmental confounders that a MOX resistance otherwise folds into its gas response.

03What it produces

Accuracy that survives sensor aging.

Accuracy recovered from drift. Domain adaptation holds classification accuracy as the sensor ages, where an uncompensated model decays batch by batch.

A calibration model that holds as the hardware ages: on the UCI Gas Sensor Array Drift benchmark, knowledge-distillation domain adaptation recovers up to 18 percent of the classification accuracy that drift otherwise strips away.

04The agentic loop

Closing the calibration loop.

Agentriq runs this as a closed agent loop: the model watches live sensor streams, detects when drift has moved the decision boundary, retrains or transfers calibration, and verifies before the corrected model goes back on the wire.

From work to a verified decision. The same closed agent loop: observe, verify, root-cause, judge, and write the outcome back.

01Observe: pull live sensor streams with paired temperature and humidity, track the 24-hour ambient baseline, and fit the drift-compensation correction.
02Verify the corrected output against held-out reference gases and the trained feature envelope before anything ships.
03On failure, root-cause with Agentic RAG across datasheets, drift datasets, and prior calibration runs to localize the cause.
04A judge weighs the candidate calibration against accuracy, stability, and cross-sensitivity limits, then decides accept or iterate.
05Write the accepted model and its evidence back to the device registry, closing the loop for the next reading.

05Tooling

Simulation, solvers & frameworks.

Get started

From engineering problems to autonomous solutions.

Talk to our engineers john.ferderigos@agentriq-sa.com