Live status / Weather
Weather
Radar forecasts, natural-colour satellite pictures, thermal sharpening, storm features and a forecast for the ship that loses its shore data. These cards give the research position as of 21 October 2026; the live lines and counters read the latest job feed.
-
Radar nowcast: v3.1 serves; v3.3 passes skill, misses sharpness
In plain words: we take the radar picture we have right now and predict where the rain will be in thirty and sixty minutes, and we keep a version only when it beats the standard method on days it has never seen.
Decision openThe radar nowcast carries the latest rain field forward. Version 3.1 serves on the weather site. Version 3.3 moves the rain skilfully but still smooths away too much fine detail.
Fractions skill score measures how closely forecast rain coverage matches radar within a local area. Version 3.3 clears its bars at both forecast leads; sharpness remains below the required level. The open decision is how to weigh skill against sharpness. No job for this line is on the supercomputer now.
Skill clears the dashed bars at both leads; sharpness stays below its bar. -
Six-hour nowcast: collecting data for the research line
In plain words: the same idea stretched to six hours ahead, which needs satellite and weather-model fields as well as radar; today we are collecting those inputs.
CollectingThis research line would extend the radar forecast to six hours by blending its motion with weather forecasts and satellite imagery.
Partner data collection continues. No training run is on the supercomputer now.
Radar, satellite imagery and weather fields must meet in the collection before training. Collection calendar
Elapsed collection time from 12 October; the calendar does not verify data completeness.
-
True colour: cloud improves when the anchors go; water loses its gain
In plain words: a satellite does not see colour the way an eye does. We teach a model to turn its red, blue and vegetation bands into the picture a person would see, and the hard part is keeping the sea the right colour at sunrise and sunset.
v7e queuedThe v7 family learns a correction to a physical colour recipe. A time-matched reference teaches the correction only where the satellite scans agree in time. Version 7c brought low-sun ocean error down to 12.73 against the recipe’s 16.87, but cloud error rose.
The 20 October tests separate the causes. Their class anchors assign a fixed reference colour to cloud, land or deep ocean. Removing the cloud and land anchors improves cloud; removing the deep-ocean anchor loses the water gain. The anchors had carried the water correction.
Both v7d arms bring the pooled error to 10.83 and 10.49, with every seasonal test day under 12 for the first time. They still miss the low-sun checks. Version 7e keeps the anchor only on unmatched deep ocean. Its final training piece is queued behind a full lane; v3 still serves.
Paired bars show water improving in v7c, then cloud improving as water worsens in v7d. -
Sharpening: freezing the trunk keeps the thermal gain
In plain words: the satellite's heat band sees the world in 2 km blocks, its visible band in half-kilometre blocks. We teach a model to bring the heat band to the finer grid, day and night, without inventing detail that is not there.
Product decisionThe thermal band C13 is being sharpened from 2 km to half a kilometre using visible imagery as a guide. Tests at reduced scale have measured truth; the native-scale look has no measured half-kilometre thermal reference.
The full-resolution training phase had eroded the gain. On 20 October, freezing the trunk (the main feature extractor) and training only the final block kept +7.14 dB over bicubic enlargement with no erosion. Letting the trunk change had fallen to +1.7 dB at the selected checkpoint. W0 stays at reduced scale; W1 lets the trunk change; arm C freezes it.
Seeds 43 and 44 repeat the result at +7.15 dB each, a spread of 0.012 dB across the three seeds. The training recipe is settled. The product decision remains open: whether the native-scale pictures look plausible enough to serve.
W1 loses gain after the phase change; W0 holds, and C’s reported range stays close to it. -
Radar features: the corpus is adopted; chip building is under way
In plain words: a forecaster can spot a hook echo or a rotating storm in Doppler radar by eye. We are teaching a model to do the same, so the site can draw those signs as a layer the moment they appear.
Building chipsOpened on 12 October, this line aims to mark storm features in Doppler radar: hooks, rotation, hail and damaging-wind signatures. The training corpus is now adopted on the supercomputer: 75 of 85 months of Level II radar, totalling 2.18 TB, plus 152 GB of TorNet tornado cases.
The chip builder cuts labelled storm windows from those volumes. Twelve shards have been working on one system since 03:02 UTC on 20 October, with pieces waiting when its lane is full; twelve more are on a second system’s allocated lane. Next come merging the shards, staging the chips, and comparing three sizes at three seeds each.
Serving, medium and large candidates will use the same chips so size can be compared. -
The data-denied ship: control complete; imagery is filling the gap
In plain words: a ship can lose its data feed and keep only the satellite pictures. We are training a forecast model that starts from the last state the ship had and corrects it from imagery alone, so the forecast stays useful for days instead of hours.
Collecting imageryWhen a ship loses its shore data, can satellite pictures keep a global forecast useful? The full-data control is complete. It starts the same forecast from an official atmospheric analysis, providing the reference against which the imagery experiment will be scored.
The learned analysis starts from the last state the ship had, aged 24 hours. It corrects that state using 48 hours of infrared and visible imagery on a one-degree grid, with a new cycle every six hours. The forecast then runs from that corrected state.
ERA5 is complete. Imagery is complete for June 2017 to October 2018 and July 2023 to December 2025. The October 2018 to June 2023 gap is being fetched and reduced twelve months at a time, about 290 GB raw per month. The current pieces launched at 02:49 UTC on 21 October. Next come the three training seeds.
The archive’s ends are complete; the grey middle is filling, starting again at October 2018. -
Forecast writing
TrainingA language model that writes and discusses weather and ocean forecasts from numerical model output, scored against what the weather did. It is part of the weather study and keeps its own page with its stages.