Simba, the End-to-End AI Meteorological Model With Billions of Parameters

TerraQuanta acquires large-scale weather observation data from tens of thousands of stations from around the world, to enable AI to automatically learn climate change patterns and deliver weather forecasting with higher accuracy.

End-to-End AI

Simba is a transformer-based, 5-billion-parameter AI model. Unlike conventional weather patterns like WRF, Simba is completely based on AI and huge amount of data.

Fast, 50000 Times Faster

Simba's forecasting efficiency is 50000 times higher than conventional methods. This not only saves power, but also makes it possible to deliver large-scale ensemble forecasting.

Competitive Results

Simba has achieved competitive accuracy compared with meterological agencies in 15-day forecasts. It has been validated to be an effective source of weather data for power forecasting.

On Climate Change And Extreme Weather Conditions

Regular Re-Training and Optimization of Models

  • There is a potential risk of significant climate change in the following couple of decades

  • Weather observation data is updated quarterly in order to re-train our models. Climate changes, so do the models.

  • With historical weather data of more than 40 years, Simba is an expert of predicting extreme weather conditions

Optimization Towards Wind Field and Solar Radiation

Simba makes Wind Field And Cloud Movements Predictable

  • Wind field and cloud movements are essential for wind farms and solar power plants, respectively

  • And they are both hard to be predicted

  • In combination with other technologies such as remote sensing, Simba provides reliable results for wind field and cloud movements predictions.

*10-m Wind filed forecast by Simba and the observed data(reanalysis)

Simba Meets Satellite AI

The more data used, the better results expected. That's how AI models work

  • Geostationary-orbit satellites are used for data assimilation at high frequency.

  • The data from geostationary-orbit satellites is updated every 10-15 minutes, used to train and validate models

  • We are also running data fusion to put analyzed data, re-analyzed data, as well as observed data together to train AI models.