We are pleased to announce a new funding award from the Navy’s Small Business Innovation Research (SBIR) Phase I program. The project, titled “Observation Driven Multi-Scale Medium Range ML Weather Forecasts”, is supported by the Office of Naval Research to produce observation driven AI weather forecasts. 

Medium range AI weather forecasts

AI weather forecasts are increasingly showing skill across short and medium range lead times, ranging from minutes to days. Models including GNNs, Gencast, Graphcast, FourcastNet, Pangu-weather, etc. have demonstrated the ability to perform forecasts up to 15 days long. The WeatherBench project maintains a standardized evaluation benchmark that is constantly updated to compare state of the art models. Recently, the European Center for Medium Range Forecasting (ECMWF) was the first forecasting agency to release an AI driven forecast, AIFS. 

This first generation of AI weather forecasts are limited by their training data which generally includes the ERA5 reanalysis dataset. By training on reanalysis data, a gap-filled data assimilated initial state is a dependency to perform inference and produce a forecast.  However, data assimilation is an expensive process, taking hours to produce new initial conditions, with limited data ingestion capabilities.  As an increasing amount of observations are collected by government and commercial systems, the capability of high resolution data ingestion is key to improving weather forecasts. 

Direct Observation Prediction (DOP)

Rather than depending on traditional data assimilation systems, DOP aims to generate AI weather forecasts from only observations. Our work on EarthNet has been a step in this direction — producing gap-filled initial conditions by conditioning a multi-modal generative model on a diverse set of observations. This progress provides a base for Zeus AI to develop a suite of forecasting tools that are independent of government models. As such, this SBIR award will develop a medium range AI forecast (1-15 days) model on EarthNet initial conditions. This model will provide the following benefits beyond the first generation of medium range AI weather forecasts:

1. 1. Learn from observations
   2. Low latency 
   3. High-resolution 
   4. Multi-initial condition ensembling

The Navy SBIR program provides non-dilutive funding for early-stage research and development with significant commercialization potential. The support by the Office of Naval Research is appreciated and over the next year we look forward to sharing the project’s progress.