IARC’s Climate Physics Journal Club is starting a seminar series on the use of Artificial Intelligence and Machine Learning in scientific applications. Contributions are invited from all fields of science. The series will launch with climate-related applications developed by scientists from the Los Alamos National Lab.

Circumpolar observations of ice wedge melting and thermokarst pool expansion

Presenter: Chuck Aboldt, Director’s Postdoc at Los Alamos National Lab, Earth and Environmental Sciences Division

Abstract: Recent permafrost thaw has progressed at rates that far exceed predictions from Earth system models (ESMs), which are the primary tools used to project interactions between the global carbon cycle and climate. Accurate simulations of permafrost degradation are vital to forecasting climate change, because vast stocks of carbon may be mobilized as frozen sediments thaw. One of the key sources of error in pan-Arctic scale simulations of permafrost thaw is the absence of a fine-scale, but ubiquitous process: the melting of ice wedges, or meter-scale bodies of ice buried at the top of the permafrost. Ice wedge melting accelerates thaw, because it tends to create small ponds, or thermokarst pools, which absorb heat more efficiently than the tundra they replace. Building this process into ESMs is challenging, due to incomplete knowledge of the highly variable extent to which ice wedges across the Arctic have responded to climate change thus far. To address this challenge, my research analyzes an enormous dataset of submeter-resolution satellite imagery, available from 2008-present. My workflow, based on a convolutional neural network, measures the area of individual ponds and tracks their growth at twenty-five landscapes spanning the Arctic. The results reveal highly uneven decadal-scale trends, which vary from no expansion at some sites in Siberia, to a nearly fourfold increase in the area of thermokarst pools near Prudhoe Bay, Alaska. These results comprise a unique and geographically extensive dataset, which will permit parameterization and validation of the pan-Arctic simulations of ice wedge melting. The incorporation of this key process into Earth system models will improve the realism of simulated permafrost thaw, enabling more accurate projections of changes of climate change and carbon cycling over the 21^st century.

Behind the curtain: the messiness of machine learning

Presenter: Jon Schwenk, Scientist at Los Alamos National Lab, Earth and Environmental Sciences Division

Abstract: While many machine learning applications produce seemingly magical results, we often neglect the messy details of data provenance, acquisition, and curation that can account for more than 75% of the effort required to build a machine learned model. Large training datasets are often required to train robust models, which can result in focusing on data quantity over quality. Satellite imagery and derived products often serve as a primary source of this data, but can pose major technical challenges that are often overlooked. Here I will present two data-driven machine learning modeling efforts. The first is an attempt to model riverbank erosion rates globally using primarily watershed-averaged characteristics. I will show a tool we developed (RaBPro) to aid in generating spatially-relevant predictors and some preliminary modeling results. The second project aims to fuse millions of in-situ ocean observations with MODIS satellite imagery in order to create a temporally- and spatially-continuous dataset of sea surface temperatures, salinities, and turbidities that will aid Earth System Model development. I will highlight some of the technical challenges of both efforts as well as present (some of) our solutions to overcome them. I will close by presenting an InteRFACE-generated dataset that could serve as a collaborative machine-learning project.

Watch the recording [Passcode: AJGBiM?6]

About the Climate physics journal club: Exchanging ideas and findings in an informal atmosphere– scientific discussion welcome!