Gains:
- Ability to distinguish between emissions (from the source) and concentration (measured in air) and convert the calibrated data to AQI and compare it with thresholds
- Ability to build a short-term pollution prediction model and present each prediction honestly with confidence intervals and model limits
- Ability to support public health claims with evidence without having artificial intelligence adapt parts that require distribution/physics
Air pollution is invisible but deadly: it causes millions of premature deaths every year and is intertwined with climate change. Institutions are responsible for both their own stack emissions and the air quality around them. In this unit, you will learn air quality parameters, measurement and modeling methods, prediction with AI and the limits of all these.
Concepts first. Air quality is generally measured on several major pollutants: PM2.5 and PM10 (particulate matter smaller than 2.5 and 10 microns — respirable dust), NO₂ (nitrogen dioxide — from combustion), SO₂ (sulfur dioxide), O₃ (ground level ozone), and CO (carbon monoxide). AQI (Air Quality Index) is the index that converts these pollutants into a single colored/numbered scale (good-dangerous) that the public can understand. Emission is the amount released from the source, while concentration is the concentration in the air; It is critical not to confuse the two.
Hint: Emission (from the source) and concentration (measured in air) are two different things. Wind, temperature and topography cause the same emission to result in very different concentrations. Saying "there is less coming out of the chimney" does not mean "the air nearby is clean".
Modeling: from emission to concentration
Calculating the concentration of a pollutant coming out of a chimney will reach a nearby neighborhood is a matter of dispersion modeling. These models use meteorology (wind direction/speed, atmospheric stability) and topography. Classical models (e.g. Gaussian distribution model) are physics-based; AI is used to accelerate these models, interpret the output, and extract patterns from large measurement data sets — but it is dangerous to “make” physics fit into AI.
Forecast: predicting tomorrow's weather
AI's strongest field is forecasting: building models that predict pollutant levels tomorrow or a few hours later, using historical measurement, meteorology and traffic data. This is valuable for public health warnings and precaution (e.g. traffic restriction). But every prediction has uncertainty, and the model will be wrong when it falls outside the conditions it was trained on (for example, an unexpected fire).
Step by step: an air quality analysis
1. Select the question and parameter. Which pollutant, which location, for what duration?
2. Verify the data source. Reference station or low-cost sensor? Caliber?
3. Pre-process. Clean up missing data, sensor drift, extreme outliers.
4. Convert to AQI and compare with threshold. Is the legal limit exceeded?
5. Model/predict. Trend, seasonality, future projection.
6. Report uncertainty. Always present estimates with confidence intervals and model limits.
three mini cases
Case 1 — Emission-concentration confusion. One facility wanted to announce that it had reduced its stack emissions by 20% and therefore "neighborhood air was 20% cleaner". The expert showed that the wind direction changed at that time and the concentration was actually affected by meteorology; The emission reduction was real, but the concentration claim was unsubstantiated. The statement was limited to "we reduced stack emissions by 20%"; Air claim without evidence removed.
Case 2 — Sensor calibration. A school received a "dangerous PM2.5" alert with a low-cost sensor. When the expert compared the sensor to a nearby reference station, he found that the value was inflated at high humidity; After humidity correction, the value decreased to the "medium" level. An uncalibrated sensor could cause panic. (This once again shows how critical sensor calibration is in every measurement area.)
Case 3 — Forecast uncertainty. A municipality warned of "high pollution" the next day with its AI forecast model, but an unexpected sandstorm foiled the forecast. The team had previously reported that the model was “unreliable for non-training data events”; therefore the warning was presented as "possible" and there was no loss of confidence. Making uncertainty clear is the responsible use of prediction.
Weak prompt / Strong prompt
Weak prompt:
Predict tomorrow with this weather data.
Why it's weak: No demand for parameters, location, data quality, model type and uncertainty. The output would be a single number, its reliability unknown.
Powerful prompt:
Your role: air quality data scientist. With the following PM2.5 hourly data and meteorology (wind, humidity, temperature), set up a simple forecast model (e.g. gradient boosting) in Python for the next 24 hours. Steps: (1) data cleaning and sensor drift control, (2) feature preparation, (3) model, (4) give predictions with confidence interval. Clearly state the limitations of the model (unreliability in non-educational situations). Data: [here]
Four copyable templates
1) Air data cleaning + AQI:
Clean up the following air quality measurements with Python (undershoot, stuck, drift, overshoot) and convert them to standard AQI. Specify which AQI formula you use; fitting. Mark hours that exceed the legal threshold. Data: [here]
2) Emission vs concentration distinction:
Check the following claim: is it about emissions (from the source) or concentration (measured in air)? Highlight the parts that confuse the two or ignore meteorology and suggest a measured restatement based on evidence. Claim: [here]
3) Prediction model (with uncertainty):
Your role: data scientist. Build a short-term forecast model (Python) for [pollutant] with the following time series. Give each prediction with its confidence interval. Write clearly under what conditions (non-training event, extreme weather) the model would be unreliable. Overly optimistic truth claim FAKE.Data: [here]
4) Resource contribution summary:
From the air quality and wind data below, analyze which direction the wind is coming from during high pollution hours to give a clue about the possible source direction. Don't say "this is the source" outright; Offer probability and verification. Data: [here]
Common mistakes
- Confusing emission with concentration. A decrease in the chimney does not mean cleanliness in the air.
- Relying on an uncalibrated sensor. Humidity and drift inflate the values.
- Presenting the estimate without uncertainty. Every estimate should be given with a confidence interval.
- Making physics fit with AI. The distribution model requires physics; The "prediction" of the LLM is not physics.
- Remembering legal thresholds. Take the limit values from the legislation.
Caution: Air quality claims concern public health. A false "clean" claim puts people at risk; An unfounded "dirty" alarm creates unnecessary panic and economic damage. Support each claim with calibrated data, appropriate model, and clear uncertainty.
In summary
Air quality; It requires separating emission from concentration, calibrating measurement, appropriate modeling and prediction presented with uncertainty. AI; It is powerful in data cleaning, pattern extraction and short-term forecasting. But physics-based distribution, sensor calibration, legal thresholds, and honest reporting of uncertainty belong to the expert. The number on the chimney is not the same as the air in the neighborhood.
Application task
Prepare hypothetical hourly PM2.5 and meteorological data for a location. With the 1st template, have the AI clean the data, convert it to AQI, and mark threshold exceedances. Then build a short-term forecast model with the 3rd template and make sure that each forecast comes with a confidence interval. Finally, have your air quality claim audited for emissions/concentration with template 2.
checklist
- [ ] I made the distinction between emission and concentration correctly.
- [ ] I checked the sensor data for calibration and drift.
- [ ] I got the AQI and thresholds from the correct formula/legislation.
- [ ] I presented the estimates with confidence interval and model limit.
- [ ] I did not have the AI adapt the parts that required distribution/physics.
- [ ] I supported public health claims with evidence.