Towards Using Neural Networks for Geoscientific Discovery
Benjamin Toms, Colorado State University
Machine learning has become exceedingly popular in studies of the Earth system. Examples of popular topics that machine learning aims to address are: Can we replace complex, computationally expensive representations of physics in climate models (i.e. parameterizations) with efficient machine-learning models? Or, can machine learning-driven forecasts supplement or even entirely replace physics-based weather and climate forecasts? Scientist should aim to understand how and why machine-learning models make their decisions when acting as emulators of the Earth system, especially in decision-driving applications such as weather and climate forecasting. If these highly parameterized depictions of weather and climate are to be trusted, then they should accurately represent the understood physics of the processes they are trained to represent. To this end, I share some results from both my doctoral studies and the field of weather and climate at large that support the ability to understand machine-learning models' decision-making process. Growing evidence suggests that interpreting the decision-making process of machine-learning models in weather and climate applications can improve their trustworthiness and reliability, aid in tuning and optimization, and even lead to discoveries of previously unknown weather and climate patterns. I will touch on each of these topics briefly, and the presentation's core will focus on using interpretable machine-learning methods to discover previously unknown patterns of Earth system variability.
Abstract Author(s): Benjamin Toms