Air-Parcel Tracks in Blocking Events
Purpose. A block is a large high-pressure system in the middle and upper troposphere that stays in place for days to weeks and diverts the usual west-to-east procession of weather around itself. This page follows the air that moves through five blocks, from just above the ground to about 200 hPa (up to roughly 12 km above sea level). Choose an event, a time in its life cycle and a start level, then click on the map: a neural network continues the paths of the 16 tracked air parcels nearest that point for 24 h into the future (forecast) and 24 h into the past (backcast), as a set of possible paths.
Terms. An air parcel is a small volume of air followed as it moves with the wind, in three dimensions; its track is the sequence of its positions. The true tracks of each event are computed by the trajectory model LAGRANTO from the hourly winds of ERA5, the European reanalysis (a reconstruction of past weather from observations and a weather model). The network is a conditional Kinematic Interpolant (Jiang, Marzouk, Bose & Pooladian, 2026): it is never given a wind field; instead it predicts how a parcel accelerates from how 192 other parcels of the same group moved during the 48 h before (for a forecast) or after (for a backcast) the chosen time. Adding small random kicks to the predicted motion gives an ensemble of 16 possible tracks per parcel; their scatter, the spread, shows how uncertain the model is. Pressure is given in hectopascals (hPa); lower pressure means higher altitude.
How it works
Parcels for every place and time. The network needs observed tracks around the point, so each event comes with a dense set of true tracks: groups (patches) of 480 parcels filling a disk of 300 km radius, released every 12 h from four days before the event to its end, on centres about 1000 km apart over the event region and up to 40° of longitude upstream (west), at six start levels: 0.967, 0.85, 0.477, 0.38, 0.285 and 0.2 of the surface pressure (just above the ground and about 850, 477, 380, 285 and 200 hPa where the surface pressure is 1000 hPa; the parcels then move up and down with the air). Each patch is followed for 156 h. In total 27,744 patches, 13.3 million parcels.
A launch. For the forecast, among the patches of the chosen level released 48–132 h before the chosen time, the one whose 16 parcels nearest the point are closest to it is used (among the six patches whose centres are nearest the point); these 16 parcels are launched, and 192 of its other parcels, observed every 3 h over the 48 h before the chosen time, are what the network sees. The backcast does the same with a patch released 24–108 h before the chosen time, observed over the 48 h after it, so the two directions may use different parcels near the point. The network then integrates position and velocity in steps of 10 minutes. A point on the map is rarely where a tracked parcel is at that moment, which is why the launch uses the nearest tracked parcels: their positions and velocities at the chosen time are known from LAGRANTO. The line under the figures gives how far from the point they start. Only the network’s tracks are drawn.
Training. The network has 32.2 million weights. An encoder (a Perceiver: 96 learned vectors of width 512 that read all observed positions and velocities by attention) condenses the observed tracks of a patch; a decoder takes the time, position and velocity of a parcel, reads the condensed tracks and returns the parcel’s acceleration and its velocity score, the direction back towards a likely track, which the Kinematic Interpolant needs for its random kicks. Each training example is a random patch at a random time, in a random direction: the tracks of 192 of its parcels observed every 3 h over the 48 h before (forecast) or after (backcast) that time, and 256 other parcels of the patch at random hours up to 24 h later (or earlier), each shifted off its true track by a small random amount (about 5 km horizontally). The network learns, by least squares, the acceleration and score of each of these parcels given by its true track and its shift. Training ran for 200,000 steps of 64 patches on two GPUs (learning rate 0.0003 after 2,000 warm-up steps, with an exponential moving average of the weights); the weights kept are those of step 158,000, which had the smallest 24-h error on validation patches.
Notebooks. Three notebooks follow the method step by step: two-dimensional tracks in stirred jets on the sphere (flows, model, minibatches, training, evaluation); three-dimensional tracks in idealized life cycles of mid-latitude storms (the dynamical core Dinosaur); and ERA5, which loads the network used on this page and true tracks of the 2010 Russian block from the same public dataset, then forecasts and backcasts them against the truth and against straight motion.
Accuracy. The network was trained on 25.3 million true tracks from ERA5 releases every 5 days of 2015–2017 at 20 levels from the surface to about 100 hPa. In 2021, a year it never saw, its deterministic 24-h forecast misses the true position by a median of 335 km and its backcast by 254 km, against 523 and 535 km for parcels moving straight on at their starting velocity. The size of the random kicks was chosen on the 2015–2017 validation parcels so that the 24-h spread matches the error; there the mean of 16-member ensembles misses by a median of 275 km at 24 h, against 338 km for the forecast without kicks. Blocks and their surroundings are not the average case, so errors near a block may differ.
Speed. The network runs in your browser (onnxruntime-web, on one processor core), stored with 8-bit weights, which changes its 24-h errors by less than their scatter from one set of random kicks to another. The first launch of a visit downloads it (40 MB); nothing is stored on your computer, and it is discarded when you leave the page. A launch then takes about 29 s on a typical computer. The tracks appear hour by hour while it runs.
Events
| Event | Life cycle shown (UTC) | How the dates and region were chosen | Studies of the event |
|---|---|---|---|
| Russia, summer 2010 | 2010-07-20 00 to 2010-08-18 00 | ERA5 TempestExtremes catalogue block 881 (2010-07-22 03 UTC to 08-13 19 UTC); Z500 minus zonal mean up to 302 m near 60N 46E on 2010-08-06 | Schneidereit et al. (2012), Monthly Weather Review Matsueda (2011), Geophysical Research Letters |
| North Atlantic, summer 2019 | 2019-07-19 18 to 2019-08-09 18 | ERA5 TempestExtremes catalogue block 1312 (2019-07-21 19 UTC to 08-05 08 UTC); Z500 minus zonal mean up to 348 m near 67N 17E on 2019-07-27, then near 73N 15W | Cullather et al. (2020), Geophysical Research Letters Sousa et al. (2020), Communications Earth & Environment |
| Pacific Northwest heat dome, 2021 | 2021-06-23 00 to 2021-07-07 12 | Wikipedia heat-wave dates 2021-06-25 to 07-07; Z500 minus zonal mean above 200 m near 53N 120-130W from 06-25 to 07-03 | Neal et al. (2022), Geophysical Research Letters Oertel et al. (2023), Geophysical Research Letters |
| California wildfires, January 2025 | 2025-01-03 00 to 2025-01-20 12 | Wikipedia fire dates 2025-01-07 to 01-31 (Santa Ana windstorm 01-07 to 01-08); Z500 minus zonal mean above 300 m near 45N 130W from 01-07 to 01-16 (415 m on 01-09) | Mass et al. (2026), Weather and Forecasting Guirguis et al. (2025), npj Natural Hazards |
| Greenland high, January 2026 winter storm | 2026-01-18 00 to 2026-01-31 12 | Wikipedia storm dates 2026-01-23 to 01-27; Z500 minus zonal mean 210-250 m near 74N 15W from 01-23 to 01-25 | no peer-reviewed study yet |
Map shading: the 500-hPa height minus its average around the latitude circle at the same time (positive: ridges and blocks); contours: the 500-hPa height every 80 m. Both from ERA5 every 6 h at 1°. Studies: peer-reviewed papers on the block or the event, blocking studies first; hover over a link for the title.
Data, code and references
- ERA5 hourly winds on 37 pressure levels at 0.25° (Copernicus Climate Change Service), read from Google’s public Analysis-Ready, Cloud-Optimized ERA5 store.
- LAGRANTO trajectory model: Sprenger, M. and Wernli, H., The LAGRANTO Lagrangian analysis tool – version 2.0, Geosci. Model Dev., 8, 2569–2586, 2015, doi:10.5194/gmd-8-2569-2015.
- Kinematic Interpolants: Jiang, Marzouk, Bose & Pooladian, Learning smooth dynamics via Kinematic Interpolants, arXiv:2610.22519 (2026).
- Tracks and maps: Hugging Face dataset; the network: Hugging Face model; code: github.com/yanxingjianken/ki-lagrangian-tracks.