March 31, 2025

GEM v1 Launch Date is April 7, 2025

We are excited to introduce GEM v1, Salient's proprietary generative AI model for sub-seasonal to seasonal forecasting. The model produces 200 daily ensembles of potential weather outcomes up to 100 days in advance. Salient GEM achieves state-of-the-art performance across medium-range and S2S timescales, outperforming publicly available models on probabilistic scoring metrics (CRPS, Brier Score) for both native variables and complex weather events.

This release includes:

• Daily availability at 0800z
• Daily resolution with 200 unique ensemble members per forecast
• Physical consistency across space, time, and output variables
• Region: North America
• Variables: Temperature (mean, min, max), Precipitation, solar insolation, relative humidity, cloud cover, 10m wind speed, 10m max wind gust, 100m wind speed, mean sea level pressure, 500mb height, HDD, CDD, wind chill, and heat index
• Hindcast for the period 2020-10-16 to 2025-04-04

<aside> ❗ Geotiff output format (data_tiff, score_tiff, forecast_tiff) will be deprecated between now and model release v9.

</aside>

<aside> ❗ Bulk Download via /download/{paths} will be deprecated with the model release v9, planned for Sept. 3, 2024. It will continue to work with v8.

• forecast_timeseries replaces download.  In salientsdk you can vectorize your call to grab years of forecasts with a single line of code.  See notebook validate. </aside>

February 24, 2025

• A synchronized tooltip displays across all models on the model comparison view, allowing you to easily compare values and probability for the same location across all the selected models. This highlights the probability differences even when the maps look visually similar.

  

• Refresh the latest init date using a refresh link next to the Time Period label. This is available in case you don't see the latest initialization dates since you started a browser session.

  

• API results and UI displays using shapefiles can return user-friendly names
  Now, the column/dimension named “Location” has the geohash and there is an additional column/dimension named “Location Name” (for CSV) or “location_name” (for nc) which has the names decoded. If a name is not supplied, the subregions are named Region 0, Region 1, etc.

• Improved debiasing met station selection

  Updated station-finding logic to use newer stations with better data records. Debiasing will only use met stations with data available after 2000.

January 27, 2025

• New! The Alerts view monitors the weather events that matter to your organization based on custom thresholds using a probabilistic view of different risk levels. This view is an extension of the Risks functionality, but uses pre-defined alert definitions that are evaluated with each new forecast initialization. View a guided tour.