Cloud Physics Conundrum: Finetuning Climate Models Using Machine Learning
Nishant Yadav, a PhD student advised by Civil and Environmental Engineering Professor Auroop Ganguly, has recently won the Best Student Paper Award by 2020 ACM KDD‘s Fragile Earth (“Feed”) workshop. Organized as part of the 2020 Association for Computing Machinery (ACM) Knowledge Discovery and Data mining (KDD) conference, the workshop website is hosted by the AI For Good Foundation, while the organizers are drawn from academia and the private industry, as well as government and nonprofit research institutes. KDD is one of the most prestigious conferences in Data Science and Artificial Intelligence (AI), while the accompanying workshops, which are peer-reviewed prior to acceptance, focus on niche and potentially new areas. Titled, “Machine Learning for Robust Identification of Complex Nonlinear Dynamical Systems: Applications to Earth Systems Modeling,” the paper will be presented at the upcoming 2020 ACM KDD Feed workshop, when the award will be presented. Two papers were selected for best paper awards in this workshop out of fourteen submissions in the niche area. In addition to Yadav and Ganguly, a co-author of the paper is Dr. Sai Ravela at MIT’s Department of Earth, Atmospheric and Planetary Sciences.
The paper centers around the uncertainty plaguing climate model projections, its implications for policy-making and risk-informed decisions; and potential artificial intelligence (AI) driven solutions. Climate models are crucial to planning an adequate response to climate change. Will a coastal town soon be threatened by rising sea levels? Will an area of land be subject to changing rain patterns that no longer make it amenable to agriculture? Will a real-estate development opportunity become more prone to flooding? Such questions are playing out in every country around the globe, and when taken together, represent decisions with trillions of dollars at stake.
Earth Systems and Climate Models: A Cloudy Picture
For climate models to be useful, however, they must be accurate. Modeling Earth’s systems and how they will interact to cause future climate is a more gargantuan task than simply collecting temperature readings over time. There are many aspects of the Earth to consider: Heat absorbed by bodies of water, the heat reflected by ice and snow, changing levels of CO2 production, deforestation, and many more. Another difficult problem to model is positive feedback loops. For example, increased temperature causes permafrost to melt and release stored CO2. These newly freed greenhouse gases contribute further warming and further carbon-releasing permafrost melt, and the cycle repeats itself.
When making a climate model, scientists must build in all such processes and account for how they interact. Sometimes, the underlying physics behind certain climate processes is not yet fully understood by science, or such processes happen on a scale that they cannot be included explicitly in the climate model. In those cases of uncertainty, modelers employ parameterizations. A parameterization is a simplified version of the suspect process that is finetuned to become more accurate as the model’s predictions are compared to real-time data.
One of the most consistent areas of trouble for climate modelers is clouds. From the wispy cirrus of a pleasant summer day to the towering cumulonimbus of an approaching thunderstorm, these seemingly benign staples of our skies can have quite an impact on climate. In some cases, clouds reflect light into space, cooling the planet. In others, they trap sunbeams on Earth, bouncing them back into the planet and causing warming. Indeed, the exact physics of clouds, how they form and interact with other systems, has long puzzled researchers who have yet to piece together the full science behind them. Modelers must employ parameterizations for cloud processes in their climate models. However, given the complexity of cloud physics and their importance to climate, these parameterizations can significantly skew the final product.
Climate Models Finetuned by Machine Learning
Climate scientists have wondered if the emerging field of AI holds the key to developing better and more accurate parameterization schemes such as one for cloud processes. But which AI method is best to use? While ML including Deep Learning has spurred renewed interest in AI and continues to provide solutions in domains ranging from image or video processing, speech recognition and machine translation to self-driving cars and intelligent robots, deep challenges remain in critical areas such as uncertainty quantification (e.g., see SDS Lab’s prior work in KDD) and interpretability (e.g., see this PNAS perspective paper).
Whereas recent works in AI-driven parameterization have used methods from classical ML (e.g., Random Forests) and Deep Learning (DL), the Yadav et al. paper examines whether such methods are necessarily the most appropriate, especially when the critical requirement for translation to risk-informed policy is uncertainty quantification (UQ) in addition to point projections. The paper takes an initial step in this direction with a highly simplified model to compare (and contrast) the performance of different ML (including DL) methods with a Bayesian learning method that naturally lends itself to UQ. While definitive conclusions about the complex climate or earth system models may not be possible with this paper, the results suggest that Bayesian methods such as Gaussian Processes may need to be seriously explored in addition to standard ML methods, including state-of-the-art DL.