Google Announces Generative AI Model to Mitigate the Uncertainty in Weather Forecasting

Google Research announced a new generative artificial intelligence (AI) model on Friday which can help mitigate the uncertainty and inaccuracies in weather forecasting. The AI model is dubbed Scalable Ensemble Envelope Diffusion Sampler (SEEDS), and instead of following the traditional probabilistic model of weather forecasting, the AI model is based on denoising diffusion probabilistic models. This is not the first weather forecasting model that the tech giant is working on, as it has previously unveiled GraphCast, a model that can predict weather up to 10 days ahead, and MetNet-3, a high-resolution forecast model for a 24-hour duration.

The announcement was made by senior software engineer Lizao Li and Google Research's research scientist Rob Carver in a blog post. The team has published a paper on the generative AI model SEEDS in the Science Advances journal. As per the announcement, the AI model will innovate weather forecasting in two distinct ways — making it more accurate and bringing down the cost to predict weather.

Highlighting the two major issues in modern weather forecasting, the paper stated that right now models run something called “probabilistic forecasts”. Essentially, they focus on the initial conditions to generate a primary forecast and as the conditions progress and the weather models receive more data, the model corrects itself to generate more accurate forecast. Google says this method allows for more uncertainty in longer-duration predictions. On costs, the research team highlighted that the massive supercomputers running highly complex numerical weather models, where the predictions need to be constantly generated to get to an accurate outcome, can run a high cost.

SEEDS, as per the research paper, works on denoising diffusion probabilistic models, which was developed by Google Research. It was trained on skill-based metrics such as rank histogram, the root-mean-squared error (RMSE), and the continuous ranked probability score (CRPS). The paper claims that while the model runs a negligible computational cost, it also improves the accuracy of the initial prediction, requiring less number of forecast generation during a particular time period.

The research team also included instances of running the AI model to predict weather and found that it offered higher reliability than the Gaussian model. Highlighting the example of a geopotential trough west of Portugal, it said, “Although the Gaussian model predicts the marginal univariate distributions adequately, it fails to capture cross-field or spatial correlations. This hinders the assessment of the effects that these anomalies may have on hot air intrusions from North Africa, which can exacerbate heat waves over Europe.” As per Google Research, SEEDS is able to account for these factors to improve its prediction. The model is yet to be peer-reviewed, and depending on its viability, might be developed into a commercial model later on.