Foehn project · live research forecast
Otso 2.0
betaupdated todayWhere will the fine particles be tomorrow, down to your street?
Every morning Otso takes the 00 UTC analysis of the Copernicus Atmosphere Monitoring Service, the ECMWF weather forecast of the same hour and the fires seen by satellite, computes the transport of the air with FLEXGREEN and forecasts PM2.5 from all its sources (traffic, industry, heating, agriculture, dust, sea salt and fires) over all land, 12 hours to 5 days ahead, as an ensemble of 8 members at 0.4° (about 40 km). A generative model then draws the same forecast at 1 km, conditioned on each member, on the weather and on the transport, and the map puts it on a street map. Otso is the coarse model of the Foehn paper, fine-tuned on the live inputs of 2026; the 1 km model is trained on GHAP V1, a satellite–station PM2.5 record.
Loading the latest forecast…
Put Otso on your phone
Scan the code with your phone to open Otso there, then add it to your home screen: Add Otso to your home screen: nothing to download from a store, it opens full screen like an app, keeps the last forecast when you are offline and updates itself every morning.
ammarkheder.com/Otso
- Scan the code, or open ammarkheder.com/Otso in Safari (not in another app's browser).
- Tap Share ⎙ (the square with an arrow, at the bottom).
- Scroll and choose Add to Home Screen, then Add.
- Otso is on your home screen and opens full screen; it keeps the last forecast when offline.
- Scan the code, or open ammarkheder.com/Otso in Chrome.
- Tap the Install Otso button below the code when it appears, or the ⋮ menu → Add to Home screen / Install app.
- Confirm. Otso appears with the other apps.
- Open ammarkheder.com/Otso.
- Click the install icon at the right of the address bar, or menu → Install Otso (Firefox and Safari on Mac: keep it as a bookmark).
- Otso opens in its own window.
Your city is remembered on this device only (nothing is sent anywhere). A new forecast appears each day around 12:30 UTC. Research forecast, not an alert: the app sends no notifications.
The forecast
Search a city, a street or an address, use your position, or tap the map: the value of that square kilometre, its class, and how the days ahead look. Scroll or pinch to zoom down to the streets, drag to pan. Layers: the daily mean of the 8 members at 1 km, the number of members above 15 µg m−3 (the WHO 24-hour guideline), and the 40 km forecast every 12 hours (mean, 90th percentile, share above 25 µg m−3).
Any city, town or village: 34,000 places above 15,000 inhabitants answer instantly (GeoNames, CC BY 4.0); streets, addresses and the other places come from OpenStreetMap as you type (© OpenStreetMap contributors). You can also tap the map or use your position. The headline value is the square kilometre that contains the place; the 12-hour curve is the 0.4° cell (about 40 km) around it.
Otso shows the forecast for the place you are, and keeps it for next time. Your position stays on this device.
Line: ensemble mean of the 0.4° cell (about 40 km) every 12 h · band: 10–90 % of the 8 members · dashed: 25 µg m⁻³. The headline numbers are the 1 km daily means.
Otso (0.4°) is downscaled to 1 km by a generative model trained on GHAP V1, a satellite–station PM2.5 record covering land from 60° S to 84° N. At monitors kept apart from everything, the 1 km forecast has an 8 to 14 % lower error than the CAMS 0.4° analysis (2022 to 2026). GHAP is itself fitted to monitors (2017–2022); the skill of each day's forecast is not known in advance.
Cities
Sixty large cities are published with every forecast; type any other city, town or village in the box to add it to the table (it is computed here from the forecast layers, and kept on this device). The 1 km column is the daily mean of the selected day at the city centre, read from the 1 km layer; the other columns are the ensemble mean over the 3 × 3 cells of 0.4° around the place, and the bars show the ten leads from +12 h to +120 h. Click a row to fly to it; sort by any column.
| City | 1 km · day1 km | Tomorrow (+24 h)Tomorrow | +72 h | +120 h | Max of the 5 days5-day max | P > 25 (max) | 12 → 120 hTrend |
|---|
How good is it?
The 1 km forecast at the monitors
The 1 km model is trained on GHAP V1 (2017–2021) and checked against air-quality monitors that were never used for anything: neither for training, nor for choosing the model, nor for any adjustment. For 2022 the 1 km fields are drawn from the archived 0.4° forecast; 2023 to 2026 are hindcasts of the live chain on the archived inputs. Root-mean-square error of the daily mean at the monitor's own square kilometre, against the monitor's day, compared with the CAMS 0.4° analysis interpolated to the monitor and with a climatology.
| Year | Monitors | Otso 1 km vs CAMS 0.4° | Otso 1 km vs climatology | CAMS 0.4° vs climatology |
|---|---|---|---|---|
| 2022 | 1,495 | −8 % | −24 % | −17 % |
| 2023 | 805 | −11 % | −24 % | −14 % |
| 2024 | 757 | −12 % | −25 % | −15 % |
| 2025 | 759 | −11 % | −25 % | −16 % |
| 2026 (Jan–Oct) | 716 | −14 % | −31 % | −20 % |
- The 1 km forecast is closer to the monitors than the 40 km analysis it starts from, by 8 to 14 % every year, and the probabilistic score (CRPS) follows. The monitors are in China (CNEMC), Europe, North America and elsewhere (OpenAQ); 2022 also includes AirKorea, Japan and India.
- Read with care. GHAP, the record the 1 km model learnt from, is itself fitted to about 9,500 monitors of 2017–2022, so a monitor-pixel comparison is favourable to any model trained on it; the gain is smaller one pixel away. The forecast skill of a given day is not known in advance; the live 1 km forecast is checked every day at the monitors (below).
Live verification of the 1 km forecast at the monitors
Every day the published 1 km forecast is kept at the pixels of about 12,000 OpenAQ monitors, then compared with their daily means once these reach the OpenAQ public archive, about four days later. Every day counts and nothing is removed afterwards; the monitors are never an input of Otso.
| Forecast day | days | monitors / day | RMSE | MAE | bias | in 10–90 % |
|---|---|---|---|---|---|---|
| Loading… | ||||||
- What is compared. The forecast daily mean at the monitor's 1 km pixel (mean of the members) with the UTC daily mean of the monitor (at least 18 valid hours). Day 1 is the day the forecast is issued, day 5 four days later. Errors in µg m⁻³, pooled over all verified days; the last column (share of the measurements inside the 10–90 % range of the members) should be close to 80 % if the spread of the members is right.
- Read with care. OpenAQ gathers official and low-cost monitors, which are not harmonised, and one monitor can differ from the mean of its 1 km pixel (a busy street, a courtyard). The 1 km model learnt from GHAP, itself fitted to monitors of 2017–2022, some of which are verified here. The 1 km model changed on 3 November 2026 (LGHAP → GHAP downscaler): the dotted line.
- All numbers: history.json, rebuilt every day.
Otso 1.2 (0.4°): fine-tuned on 2026
Otso 1.2 keeps the network of the paper and learns, for six passes, from the live inputs of January to August 2026 (243 days). Scored on 6–20 September, after its training period (no day in common with its training targets), against the CAMS analysis. RMSE of the ensemble mean over land, µg m⁻³.
| Lead | Otso 1.2 | Otso 1.1 | CAMS forecast | persistence |
|---|---|---|---|---|
| +12 h | 11.1 | 13.5 | – | 18.9 |
| +24 h | 10.3 | 13.5 | 8.8 | 15.8 |
| +48 h | 11.4 | 14.5 | – | 18.2 |
| +72 h | 12.3 | 15.2 | – | 19.1 |
| +96 h | 12.8 | 15.7 | – | 19.1 |
| +120 h | 13.4 | 15.8 | 14.7 | 19.1 |
- 15 to 24 % lower error than Otso 1.1 at every lead, and a probabilistic score (CRPS) 23 to 36 % lower: the members are both closer to the truth and better spread.
- Ahead of the CAMS operational forecast from three days on: 8 % lower error over land five days ahead (13.4 vs 14.7). The CAMS forecast keeps the lead one day ahead (8.8 vs 10.3), as expected from the system that also produces the analyses used as truth.
- One reason: the CAMS system of 2026 differs from the 2017–2022 analyses the paper's model learnt from; seeing eight months of it was enough. All numbers: scores_otso12.json.
Otso 1.1: the paper's weights on live inputs
Before publishing, Otso 1.1 was run on the live inputs of 1–20 September 2026 (20 start days) and scored against the CAMS analysis of each valid time, with and without the satellite fires, next to the CAMS operational forecast and persistence (the analysis of the morning kept unchanged). RMSE of the ensemble mean, µg m⁻³, cosine-latitude weighted.
| Lead | Otso 1.1 · land | without fires | CAMS forecast | persistence | Otso 1.1 · all cells | CAMS forecast | persistence |
|---|---|---|---|---|---|---|---|
| +12 h | 14.4 | 18.1 | – | 20.1 | 8.5 | – | 11.4 |
| +24 h | 14.5 | 15.2 | 9.2 | 17.2 | 8.7 | 5.4 | 10.1 |
| +48 h | 15.6 | 16.4 | – | 19.9 | 9.4 | – | 11.8 |
| +72 h | 16.3 | 17.0 | – | 20.7 | 9.8 | – | 12.3 |
| +96 h | 16.2 | 16.8 | – | 20.3 | 9.8 | – | 12.1 |
| +120 h | 16.1 | 16.6 | 15.0 | 20.2 | 9.6 | 8.9 | 12.0 |
Live verification of the 0.4° chain, updated every day
Every morning the forecasts of the last five days are scored against the CAMS analyses that have arrived since (nothing is hidden: every day counts). Last 30 days, RMSE of the ensemble mean over land, and the mean absolute error in the 60 published cities.
| Lead | days | Otso · land | CAMS forecast | persistence | cities · Otso | cities · CAMS |
|---|---|---|---|---|---|---|
| Loading… | ||||||
- The fires matter. With the GFAS fires (injection height
injh, the variable that replaces themamiof the training), the error over land is 20 % lower at 12 h and 3–15 % lower afterwards, and the bias over fire cells is +1 instead of −3 µg m⁻³. - Skilful, but weaker than in the paper's 2022 test. Otso 1.1 beat persistence by 25 % at 12 h and 20 % at 120 h; in 2022 it beat it by 45 % and 34 %, and overtook the CAMS forecast from 48 h on. The inputs of the live chain were checked channel by channel against the archive (winds, temperatures, humidity, pressure, rain, 13 CAMS species): all in range. What changed is the CAMS system itself, whose 2026 analyses are more variable (and richer in CO, NO₂, SO₂) than the 2017–2022 analyses Otso was trained on, and the fire product (v1.4.2 detects about twice as many burning cells); hence Otso 1.2.
- All numbers: scores.json (with the 2022 figures of the paper for comparison).
What Otso is, and is not
What it is
- The coarse Foehn model of the paper, brought up to date, and drawn at 1 km. Otso 1.2 is Foehn-CAMS, the 0.4° version trained on the CAMS analyses of 2017–2022 in the protocol of Microsoft's Aurora, fine-tuned on the live inputs of January–August 2026 because the CAMS system of 2026 differs from the one it learnt from. A second model, a pointwise flow-matching downscaler trained on GHAP V1 (a 1 km satellite–station PM2.5 record), draws the 1 km fields for each of the 8 members, conditioned on the weather forecast and on the transport computed by FLEXGREEN.
- Live inputs, every morning. The 00 UTC CAMS analysis (13 species and columns), the 00 UTC ECMWF IFS forecast from the open-data service (winds, temperature, humidity, pressure, rain, to +192 h), and the fires of the previous day from GFAS. FLEXGREEN recomputes, for every land cell and every lead, where its air was.
- A probabilistic forecast. Eight members per day. At 1 km the map shows their daily mean and the number of members above 15 µg m⁻³ (the WHO 24-hour guideline); at 0.4°, every 12 hours, their mean, their 90th percentile and the share above 25 µg m⁻³; the city table and the NetCDF file also give the 10 % and 90 % quantiles.
- Where there is no 1 km value. The 1 km model covers the land where GHAP is valid, from 60° S to 84° N. Over Greenland and the other permanent ice, where GHAP is not reliable, the map shows Otso's own 0.4° forecast as it is; a few small land areas outside the model's 1 km mask are interpolated from the 1 km values around them or, with none within 50 km, show the 0.4° forecast. All are drawn with a light hatching, and the value box says which source you are looking at.
- Reproducible. Each day's inputs, their sources and the exact variables used are logged with the forecast (see the
inputsfield ofindex.json). The real-time chain is on themainbranch of the Otso repository, the model code in the Foehn repository.
What it is not, yet
- Not a street-by-street measurement. One value per square kilometre, from a model: streets, tunnels and chimneys are not resolved. The street map is there to find your way, and the 1 km value is a neighbourhood value.
- Not the 1 km Foehn of the paper. The paper's kilometre-scale forecasts need a 1 km analysis that exists only for the past; the live 1 km fields come from a downscaler of the 0.4° forecast, trained on GHAP. GHAP is itself fitted to monitors (2017–2022), which the station scores above inherit.
- Not validated in real time at 1 km. The station scores come from 2022 to 2026 hindcasts; the daily 1 km forecast has no live verification yet. The 0.4° chain is verified every morning against the CAMS analyses (scores).
- Not an operational service. It runs on a shared supercomputer (LUMI) when a GPU is free; a day can be late or missing. The last seven days of the 0.4° forecast are kept, the latest day of the 1 km tiles.
- Not an alert. See the notice at the top. Otso is named after the bear of Finnish folklore. Versions: 1.1, the paper's weights on live inputs (10 October 2026); 1.2, fine-tuned on January–August 2026; the 1 km layer since 4 November 2026. A thing of the forest, best observed from a distance.
The previous page, with the 0.4° forecast drawn on a plain world map without a street basemap, is kept as the classic map; it also works in browsers without WebGL.
Data
Everything the page shows is in two branches of the repository: live (0.4°, the last seven forecasts) and live1km (the 1 km tiles of the latest day), both rewritten every day.
| NetCDF of the day | otso_<date>.nc — mean, 10/50/90 % quantiles and P(>25) of PM2.5, 0.4°, ten leads, 0.1 µg m⁻³ precision |
| Cities | cities.json — the table above (0.4°) |
| Inputs and provenance | index.json — sources, versions, the fire-injection variable, land mean by lead |
| Maps | map_<lead>.png — the rendered map of each lead; the mean_, p90_ and prob_ PNGs are 8-bit data layers (see png in index.json) |
| 1 km tiles | tiles.json — the encoding (the palette index of each pixel is the value, the class and the source), the existing tiles, the domain; the tiles are <day>/<product>/<lead>/<z>/<x>/<y>.png (Web Mercator, 256 px, zoom 0–7); latest.json names the day |
| Branches | github.com/AmmarKheder/Otso/tree/live (latest.json lists the available days) · tree/live1km |
| Basemap | OpenFreeMap, © OpenMapTiles, data © OpenStreetMap contributors. Land and lakes of the 1 km tiles: Natural Earth. Place and street search: GeoNames (CC BY 4.0) and OpenStreetMap data through Photon (komoot). |
| Sources | Contains modified Copernicus Atmosphere Monitoring Service information 2026 (CAMS analysis, GFAS). ECMWF open data, CC BY 4.0. GHAP V1 (Wei et al. 2023). Neither the European Commission nor ECMWF is responsible for any use of this forecast. |
| Citing | Please cite the Foehn paper (see the main page) and name the day's forecast by its issue date. |