Global calibration for CO and Ozone on the Multi-Gas Module

Clarity's Multi-Gas Module measures NO₂, Ozone (O₃), and CO with electrochemical sensors. Like all electrochemical air sensors, they respond to temperature and humidity as well as to the pollutant itself, so their raw readings need calibration. Clarity now provides global calibration models for CO and Ozone, joining the global calibrations that already exist for PM2.5 and NO₂. This article explains how these models were developed, the accuracy you can expect, and their limitations.

note

This article is about the Multi-Gas Module. Clarity's dedicated Ozone Module uses FEM-grade optical detection and needs no field calibration — only a factory recalibration every two years.

What a global calibration is

A global calibration is a single model, developed from Clarity's pooled collocation data, that is applied to every device in the fleet. It converts the sensor's raw signal into a calibrated concentration using the pollutant signal plus the module's own temperature and humidity measurements.

Because the model is pre-built, no site-specific collocation is required: your Multi-Gas Modules produce calibrated CO and Ozone data from the moment they're deployed. That's the central benefit. It removes the four-plus weeks of collocation setup, logistics, and analysis that a collocation-based calibration requires, so projects can start reporting data immediately, including in regions with no reference monitor to collocate against.

The trade-off is that one model must serve every device, at every site, in every season. The sections below quantify what that trade-off looks like in practice.

How the models were developed

The CO and Ozone global models were trained on a long-term collocation of 15 Multi-Gas Modules alongside a regulatory-grade reference station in the eastern United States, spanning roughly nine months from summer through early spring, so the training data covers a wide range of temperatures and humidities.

The models were then evaluated on data they had never seen:

• New devices: additional modules at the training site that were excluded from model development, to measure how well the model transfers to devices it wasn't trained on.
• New sites: independent collocations at eight other reference sites across North America (including a high-elevation site running for over a year), none of which contributed any training data.

The Ozone model corrects for temperature and humidity (via dewpoint); the CO sensor's response turned out to be more stable against weather, so its model is a simple gain-and-offset correction. All modeling and evaluation is done on hourly averages.

Expected accuracy

The numbers below are median results from the held-out evaluations described above. They describe a typical device deployed at a site the model was never trained on.

	Ozone	CO
Correlation with reference (R²)	typically 0.85–0.95 (full range 0.73–0.96 across evaluation sites)	typically 0.80–0.85 (full range 0.60–0.85)
Typical hourly error (RMSE)	~7–10 ppb	~0.06–0.17 ppm
Typical site-level bias	within about ±6 ppb	typically ~0.05 ppm; offsets of −0.35 to +0.16 ppm observed
Device-to-device spread	95% of devices within about ±5 ppb of the fleet response	95% of devices within about ±0.04 ppm

At the training site itself the Ozone model performs close to the US EPA's air-sensor performance targets for O₃ (median hourly RMSE under 5 ppb, R² ≈ 0.9). At new sites, the additional site-level bias means globally calibrated data should be treated as indicative: excellent for tracking patterns, trends, hotspots, and relative differences, but not equivalent to a reference monitor or to a collocation-calibrated sensor.

Benefits

• Deploy immediately. No collocation period, reference-site access, or per-project modeling before you get calibrated data.
• Works where collocation isn't feasible. Many monitoring projects are in places with no accessible reference station. A global model is what makes calibrated CO and Ozone data possible there at all.
• Strong correlation with reference. The models track hourly and daily pollution patterns well (high R² at almost every evaluation site), with low-to-moderate absolute error.
• Consistent methodology. Every device runs the same documented model, so measurements are comparable across your fleet and over time.

Limitations

We believe transparency about what a global model can't do is as important as what it can.

• Device-to-device baseline offsets. Each electrochemical cell has its own zero offset, and a global model, by construction, corrects the average device, not your specific one. This is the largest remaining error source: individual devices can sit a few ppb (Ozone) or a few hundredths of a ppm (CO) above or below the reference, as a roughly constant offset. For CO in particular, nearly all of the observed transfer error is this kind of flat baseline offset rather than a distorted response.
• A limited (but growing) collocation dataset. These first-generation models were trained at a single reference site and evaluated at a small number of independent sites. Conditions outside what that dataset covers (for example, extremely hot-humid or hot-dry climates) are less validated, and performance there is less certain.
• Ozone cross-sensitivity to NO₂. The electrochemical Ozone measurement is derived in combination with the module's NO₂ cell, so ambient NO₂ can influence the Ozone reading. In environments with high NO₂ concentrations (for example, heavily trafficked urban locations), Ozone accuracy may differ from the expectations above. Characterizing this interaction is an active part of our ongoing research.
• Sensor drift. Electrochemical cells lose sensitivity over time. A static global model cannot correct drift, so accuracy expectations apply to the first year of deployment, and long-running projects should plan for periodic performance checks or recalibration.
• Not a substitute for collocation where accuracy is critical. For projects that need the highest accuracy, or that need documented, device-specific performance metrics, a collocation-based calibration remains the recommended approach. It corrects each device's individual baseline and tailors the model to local conditions.

Ongoing research

These are v1 models, and we're continuing to develop them as our collocation dataset grows. Active research directions include per-device baseline estimation in the field (which directly targets the largest error source above), improved handling of the Ozone–NO₂ cross-sensitivity, drift detection and management, and corrections for high-elevation deployments. As new collocations mature, we'll validate the models across a wider range of climates and update this article as the models improve.

What's next

• See calibration details by pollutant.
• Check which calibration is applied to your devices.
• Learn how to collocate your Clarity device with a reference monitor.
• Understand R² and RMSE, the metrics used above.