The Architecture Behind DeepMind’s Model for Near Real Time Weather Forecasts

Deep Generative Model of Rain (DGMR) is the newest creation from DeepMind which can predict precipitation in short term intervals.



I recently started a new newsletter focus on AI education and already has over 50,000 subscribers. TheSequence is a no-BS( meaning no hype, no news etc) AI-focused newsletter that takes 5 minutes to read. The goal is to keep you up to date with machine learning projects, research papers and concepts. Please give it a try by subscribing below:



Last week, DeepMind once again made news by releasing another ground breaking deep learning research. In a paper published in Nature, DeepMind proposed a model for 1–2 hour precipitation forecasting (nowcasting). Today, I would like to discuss some aspects about the architecture used by the DeepMind models.

Weather forecasting is one of the classic problems for ML techniques almost since its inception. While today we have powerful forecasting models that are able to accurate predict weather conditions several days in advance, most of them struggle to short-term forecasts. This type of forecast is incredibly important in domains such as agriculture or transportations in which short-term weather changes can have significant impacts. To address the challenges in nowcasting scenarios, DeepMind relied on a very traditional deep learning technique.


Generative Models for Nowcasting

From a functional standpoint, DeepMind model focuses on making predictions based on radar images. More specifically, the model predict radar images based on radar images. For this scenario, DeepMind used a deep learning method known as deep generative models(DGMs). To make things simpler, DeepMind called its model Deep Generative Model of Rain (DGMR).

Image Credit: DeepMind


DGMs are statistical models that learn probability distributions of data and can generate new samples from their learned distributions. DGMs are specially well suited for weather forecasts as they are not only able to learn from specific distributions but to represent uncertainly in both spatial and temporal settings.

DeepMind’ss DGM is based on a generator trained using on two discrimination networks and an additional regularization layer. The two discriminators focused on temporal and spatial data respectively. Both components are based on generative adversarial networks(GANs) with semi-identical structures. The regularization layer includes special attention blocks designed to prevent overfitting.

Image Credit: DeepMind


DeepMind evaluated its generative model for medium and heavy rain forecasts and presented the results to a large number of meteorologists who preferred DGMR to alternative forecasts. Below you can see a comparison between DGMR’s forecasts and alternative approaches advection approach (PySTEPS) and deterministic deep learning methods (UNet).

Image Credit: DeepMind


DGMR represents a relevant evolution in nowcasting scenarios and a premier examples of how deep learning methods can help in these domains. Some of the techniques used in DGMR are likely to inspire new models in this challenging but important domain.

Bio: Jesus Rodriguez is currently a CTO at Intotheblock. He is a technology expert, executive investor and startup advisor. Jesus founded Tellago, an award winning software development firm focused helping companies become great software organizations by leveraging new enterprise software trends.

Original. Reposted with permission.