How the estimates are made, and checked

The grid has 348 cells, each 0.1° (about 10 km) on a side, covering 36.75–39.57°N and 76.76–75.69°W: the Bay from Hampton Roads to the Susquehanna. Everything is hourly in UTC (Coordinated Universal Time, 4–5 hours ahead of Eastern time). Cell IDs are R{row}_C{column}, counted from the south-west corner.

Data splits are by calendar date: train 2018-01-01 to 2024-12-31, validate 2025, test 2026-01-01 to 2026-03-31. Nothing from the test period is used for training or for choosing settings.

NO₂ (µg/m³, surface). Pandora spectrometers (in 5 grid cells) measure the total NO₂ column overhead. We convert that to a surface concentration by spreading it over the boundary layer, using an ERA5 boundary-layer-height climatology (its typical depth) by month, hour of day and location instead of a fixed 1 km. Where an EPA surface NO₂ monitor lies within 0.30° of a cell centre (14 monitors, 11 cells), its real reading replaces the converted value.

O₃ (µg/m³, surface). Built from NASA TEMPO's tropospheric ozone profile product, converted to a surface value, with real EPA surface O₃ readings replacing it within 0.30° of a monitor (33 monitors, 25 cells). Earlier versions used Pandora's total ozone column in Dobson units (a measure of all the ozone overhead), which is dominated by the high-altitude ozone layer and can't be compared with surface monitors.

The target for each hour is the next hour's value: a 24-hour window ending at hour t predicts hour t+1.

The map's LSTM uses 33 inputs per hour:

• Current NO₂ and O₃ at the cell (the target series itself, where measured).
• Neighbouring NO₂: the same hour's NO₂ at the nearest other cell with ground truth, and the distance to it. This gives unmonitored cells a live signal.
• Aerosol: AERONET optical depth (how hazy the air is), Ångström exponent (a hint at particle size) and water vapour (6 sun photometers), and VIIRS satellite aerosol optical depth.
• Weather and calendar: ERA5 boundary-layer-height climatology, hour of day, day of year, day of week and a weekend flag (as sine/cosine pairs where cyclic).
• Place: location encodings, elevation, road density, land-cover fractions and population density.

A missing input gets that feature's training-set median. Inputs and targets are rescaled with a RobustScaler (which puts every input on a similar scale without letting rare extreme hours dominate), fitted on the training years only. TEMPO NO₂, TEMPO total-column O₃ and TROPOMI NO₂ were tested and dropped: they lowered test skill.

LSTM (run o3_surface_v1): two stacked LSTM layers (224 then 112 units), each followed by layer normalisation and 14% dropout (standard techniques that keep training stable and stop the network memorising), a shared 128-unit dense layer, and separate one-number outputs for NO₂ and O₃. Input is the last 24 hours of all inputs. Trained with Adam (learning rate 4×10⁻⁴, gradient clipping) on a Huber loss (an error measure that is less thrown off by a few extreme hours).

Baselines: linear regression, random forest, gradient boosting and XGBoost, trained on the same features, targets and splits, using only the last hour of each 24-hour window. Linear, random forest and XGBoost are on the map; all four are on the validation page.

Persistence (“next hour = this hour”) and climatology (a monitor's average for that month and hour) are not trained models, but they are the bars every model has to clear.

Code: lstm_model.py, baselines.py, run_lstm_pipeline.py, o3_target_fix/run_o3_surface_pipeline.py.

Fit to target R² against the model's own target on the January–March 2026 test hours. It shows the model learned the target; it says nothing about whether the target is right.

EPA check Each EPA monitor is matched to the nearest cell centre within 0.20°; monitors sharing a cell are averaged per hour; readings in parts per billion or million are converted to µg/m³ at 25 °C and sea-level pressure. Each prediction is compared with the EPA reading for the hour it predicts (t+1), and persistence is scored on exactly the same cell-hours. R² is computed per cell, then averaged. An independent script recomputes every number from the saved prediction files.

Unmonitored check The same EPA check at a cell that has a monitor but no ground truth of its own in the target: the closest thing to a test away from instruments.

Calibration test. To see whether a statistical correction could rescue the original model, its output is rescaled per site (EPA reading = slope × prediction + intercept, fit on 2018–2025) and scored on the same January–March 2026 hours. It removes the bias but adds almost no skill, which is why the later models train on real EPA readings instead. Code: aqs_integration/calibrate_and_validate.py.

EPA check Second check, on new data. The same scoring on April 1 to October 5, 2026, beside persistence and climatology on the same hours. Certified EPA data lag by months, so most of these readings are AirNow's preliminary data for the same monitors (in late March they match the certified values exactly for O₃ and at r = 0.997 for NO₂). Sites with under a month of readings are left out of the averages and listed.

Code: o3_target_fix/aqs_validation_all_models.py, o3_target_fix/verify_aqs_validation.py, o3_target_fix/nowcast_check_apr_oct.py.

For every cell and every hour from April 1 to October 5, 2026, the models read the previous 24 hours of inputs and estimate the next hour, exactly as in training and testing. The inputs are built by the same code that built the training data, with newer readings added:

• EPA monitor readings (certified where available, otherwise AirNow preliminary) feed the NO₂ and O₃ inputs at cells with a monitor nearby, through the same blending step as in training;
• the neighbouring-cell NO₂ input comes from the nearest such cell's reading for the same hour;
• calendar features and the boundary-layer climatology come from the timestamp;
• AERONET and VIIRS, which have no data in the cache after March, get their training median, as in training whenever they were missing.

The model never sees the reading for the hour it estimates. The map shows each day's mean of these hourly estimates. See how they score.

Code: o3_target_fix/nowcast_check_apr_oct.py (--web).

The next-hour map shows each cell's daily mean on one fixed colour scale for every model and day (NO₂ 0–40, O₃ 0–120 µg/m³, viridis). For display it is smoothed between cell centres with inverse-distance weighting (nearby cells count more than distant ones); the downloadable values are the cell values. The AQI view converts EPA breakpoints to µg/m³ at 25 °C; the AQI uses 1-hour (NO₂) and 8-hour (O₃) averages, so colours on daily means are indicative only. The backtest map shows each cell's error against the training target.