生成AIによる天気予報の不確実性定量化

<span class="byline-author">Posted by Lizao (Larry) Li, Software Engineer, and Rob Carver, Research Scientist, Google Research</span>

<img src="https://blogger.googleusercontent.com/img/b/R29vZ2xl/AVvXsEglI5U51vvhkA4cAuVvMLn0TbbL5pdlFL-LO1sNnqLyUieA6A88I5HrhJlszxR1GKQqSK5wsdlATDKSy6EC1BsNF7tzS6oVlFLtau13mVFLk954nFu85HDMP3PrQboG4eXExEtUjEuDRFpcrMqE_F0ikSwXiWBECAfJiLbjr6h6523DROJkbC284xX35zC7/s1000/image3.gif" style="display: none;" />

<p>

Accurate weather forecasts can have a direct impact on people’s lives, from helping make routine decisions, like what to pack for a day’s activities, to informing urgent actions, for example, protecting people in the face of hazardous weather conditions. The importance of accurate and timely weather forecasts will only increase as the climate changes. Recognizing this, we at Google have been investing in weather and climate research to help ensure that the forecasting technology of tomorrow can meet the demand for reliable weather information. Some of our recent innovations include <a href="https://blog.research.google/2023/11/metnet-3-state-of-art-neural-weather.html">MetNet-3</a>, Google's high-resolution forecasts up to 24-hours into the future, and <a href="https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/">GraphCast</a>, a weather model that can predict weather up to 10 days ahead.

</p>

<a name='more'></a>

<p>

Weather is inherently stochastic. To quantify the uncertainty, traditional methods rely on physics-based simulation to generate an ensemble of forecasts. However, it is computationally costly to generate a large ensemble so that rare and extreme weather events can be discerned and characterized accurately.

</p>

<p>

With that in mind, we are excited to announce our latest innovation designed to accelerate progress in weather forecasting, <a href="https://www.science.org/doi/10.1126/sciadv.adk4489">Scalable Ensemble Envelope Diffusion Sampler</a> (SEEDS), recently published in <em><a href="https://www.science.org/journal/sciadv">Science Advances</a></em>. SEEDS is a generative AI model that can efficiently generate ensembles of weather forecasts <em>at scale </em>at a small fraction of the cost of traditional physics-based forecasting models. This technology opens up novel opportunities for weather and climate science, and it represents one of the first applications to weather and climate forecasting of probabilistic diffusion models, a generative AI technology behind recent advances in media generation.

</p>

<br />

<h2>The need for probabilistic forecasts: the butterfly effect</h2>

<p>

In December 1972, at the <a href="https://www.aaas.org/">American Association for the Advancement of Science</a> meeting in Washington, D.C., MIT meteorology professor <a href="https://en.wikipedia.org/wiki/Edward_Norton_Lorenz">Ed Lorenz</a> gave a talk entitled, “Does the Flap of a Butterfly's Wings in Brazil Set Off a Tornado in Texas?” which contributed to the term “<a href="https://en.wikipedia.org/wiki/Butterfly_effect">butterfly effect</a>”. He was building on his earlier, landmark 1963 paper where he examined the feasibility of “very-long-range weather prediction” and described how errors in initial conditions grow exponentially when integrated in time with numerical weather prediction models. This exponential error growth, known as chaos, results in a deterministic predictability limit that restricts the use of individual forecasts in decision making, because they do not quantify the inherent uncertainty of weather conditions. This is particularly problematic when forecasting extreme weather events, such as hurricanes, heatwaves, or floods.

</p>

<p>

Recognizing the limitations of deterministic forecasts, weather agencies around the world issue <em>probabilistic forecasts</em>. Such forecasts are based on ensembles of deterministic forecasts, each of which is generated by including synthetic noise in the initial conditions and stochasticity in the physical processes. Leveraging the fast error growth rate in weather models, the forecasts in an ensemble are purposefully different: the initial uncertainties are tuned to generate runs that are as different as possible and the stochastic processes in the weather model introduce additional differences during the model run. The error growth is mitigated by averaging all the forecasts in the ensemble and the variability in the ensemble of forecasts quantifies the uncertainty of the weather conditions.

</p>

<p>

While effective, generating these probabilistic forecasts is computationally costly. They require running highly complex numerical weather models on massive supercomputers multiple times. Consequently, many operational weathe