Valerie Diana Smykalov , Li Li, Rodrigo Vargas, Ben Bond-Lamberty

Soil CO2 efflux, the largest flux of CO2 to the atmosphere, is expected to rise globally under climate change. Its magnitude and temporal variability are highly uncertain, and daily-scale models capturing rapid changes to environmental drivers remain rare. We used a global database of soil CO2 efflux (total observations = 7,797,535 from 2002-2020) to train a deep learning model (Long Short-Term Memory, LSTM) to predict daily soil CO2 efflux in 82 sites across gradients of climate, soil type, and land cover. The model achieved a median train and test Nash Sutcliffe Efficiency (NSE) of 0.54 and 0.02, respectively, and Kling Gupta Efficiency (KGE) of 0.67 and 0.30, respectively. The model performed well (NSE > 0.5 and KGE > 0.3) at about one-third of sites, mainly in temperate mesic ecosystems (where most training sites were located) with cyclical data patterns driven by temperature. The model performed poorly at sites with little data and noncyclical temporal patterns, mostly at extreme climates including arid, Arctic/boreal, and tropical ecosystems. The model struggled to capture soil CO2 efflux pulses and peaks/troughs, highlighting the challenges of modeling extremes in time series. Our results demonstrate that LSTM models can leverage existing data to generate synthetic daily datasets, particularly for temperate mesic regions, but also underscore the challenges of learning relationships from a spatially biased dataset. To improve model performance, future data collection should prioritize 1) historically underrepresented ecosystems with variable temperature relationships; 2) conditions under extreme weather events that may become disproportionally impactful in a warming climate.