�鶹ӰԺ

The threat of climate change is one of the greatest challenges currently facing society. Given my undergraduate studies and longstanding interest in climate change, and inspired by the success of Bioinformatics, I wanted to use machine learning to shed light on climate change. Understanding climate change is an urgent challenge. Meanwhile, climate science is an extremely data-rich field, especially considering the massive amounts of simulation output from physics-driven climate models, providing a lens into the distant past and distant future. A multi-model ensemble of these climate models is used by the Intergovernmental Panel on Climate Change (IPCC) to inform their reports to the United Nations. Climate scientists are interested in ways to address the large ensemble spread (disagreement in predictions).

As it turned out, my past work on algorithms for online learning from non-stationary (time-varying) data, with access to expert predictors, proved to be very helpful in combining and robustifying the ensemble predictions. Our online learning algorithm, which updates weights over experts (climate models), while simultaneously learning the level of non-stationarity, significantly outperformed the multi-model (non-adaptive) mean, the standard technique in climate science. This provided a proof of concept that machine learning has something to offer the field of climate science; in 2010, our work received the best paper award at a NASA conference and was discussed at an Expert Meeting of the IPCC. This momentum helped launch the field of climate informatics; in 2011, I co-founded the International Workshop on Climate Informatics, which attracted climate scientists and data scientists from over 19 countries and 30 states, within its first five years alone. The workshop turns 9 years old in 2019 (October 2-4th, in Paris).

Since then, we have developed online learning algorithms that handle both spatial and temporal non-stationarity, along with multi-resolution structure, leading to improved performance for multi-model ensemble prediction. We have also demonstrated an unorthodox application of topic modeling to discover climate phenomena from data, providing a data-driven approach to defining and detecting extreme climate and weather events. We have recently shown that deep learning can be effective in hurricane track forecasting. There are a range of other problems on which machine learning can make an impact, and we encourage both machine learning researchers and climate scientists to get involved. The climate informatics endeavor is an exciting experiment in building a new interdisciplinary research field.

•   and  from our NeurIPS 2014 Tutorial, “Climate Change: Challenges for Machine Learning.”

• In 2018, the highlighted climate informatics as a key priority.

•  .

•  .

• Additional Press about our climate informatics research.

For more information and materials, including data sets, tutorials, and workshop information, please see the .