Earlier this month Silurian released the Earth API which serves state of the art forecasts powered by our physics foundation model: GFT. We’re eager to share that we’ve added 3 new physical variables to our API response: Cloud Cover, Probability of Precipitation, and Humidity. They were added just in time to make operational forecasts of the ongoing Gulf Coast winter storm. Here are the storm predictions (in particular, probability of precipitation predictions) we made on January 18th:
Silurian’s physics foundation model approach makes it possible to add new variables to our already existing model simply by ingesting the right data. What’s more, that data can come from anywhere and it doesn’t need to be weather data. If that sounds exciting to you then keep an eye out for more news!
For now we’ve upgraded GFT to handle complex cloud physics and to predict the probability of precipitation, a feat for which you would typically need multiple numerical models. You might be curious whether this data driven approach can really do better than the reliable differential equations which drive traditional forecasting. The short answer is yes! For a fuller picture check out our evaluations below or try our API today.
Similar to our last set of evaluations, we’ll be using multiple ground truth datasets from the meteorological community: weather station observations from Meteostat and analysis data from the ECMWF. The former captures ground conditions at precise locations while the latter offers a global view of the weather.
We’ll be comparing against HRES, the gold standard numerical model from ECMWF, as well as against the flagship regional models HRRR (US) and ICON (Europe). All comparisons are over the full year 2023.
We're often more interested in whether or not it's raining rather than the exact amount of rain (think about planning hiking trips, barbecues, and weddings). So this first set of evaluations computes binary classification metrics like precision, recall, and F1 scores. For each model we define a precipitation event as one where the total amount of precipitation per hour is above a standard threshold (0.1 mm). Then we compare against rain gauge data from around the world.
If you instead care about how much it's raining or snowing or hailing or sleeting or graupel-ing, we computed those metrics too! Here we display the skill score (higher is better) showing the relative improvements of GFT compared to a reference baseline.
Stations don't always report cloud cover, they're more concerned with the weather at the ground level. So, to evaluate our cloud dynamics we compare against ECMWF's IFS-Analysis which integrates a diverse set of observations into a single dataset, including the satellite data that's key to capturing cloud motion.
