The high-latitude carbon (C) cycle is a key feedback to the global climate system, yet because of system complexity and data limitations, there is currently disagreement over whether the region is a source or sink of C. Recent advances in big data analytics and computing power have popularized the use of machine learning (ML) algorithms to upscale site measurements of ecosystem processes, and in some cases forecast the response of these processes to climate change. Due to data limitations, however, ML model predictions of these processes are almost never validated with independent datasets.

In this study, we explore the appropriateness of this approach using a best-case scenario for ML model development by applying a process-rich mechanistic terrestrial ecosystem model (ecosys), to train ML models and evaluate their ability to upscale and forecast high-latitude carbon (C) fluxes under current and future climate. We first show that a ML model trained using model outputs from currently-available Alaska AmeriFLUX sites incorrectly predicts that Alaska is presently a modeled net C source. We then show that ML model performance improves substantially with increased spatial coverage of the training dataset. For example, ML model bias is halved when 240 modeled sites are used instead of 15. However, this more accurate ML model does not accurately forecast Alaska C fluxes under 21st century climate change. Whereas ecosys predicts that Alaska C sink strength will increase substantially throughout the century, the ML model incorrectly predicts that the region will remain C neutral. Using convergence cross-mapping, we show that 21st century degradation of ML model performance can be ascribed to out-of-sample changes in atmospheric CO2 concentrations, litter C inputs, and vegetation composition that cannot be captured by the ML model. Our results expose the potential for inaccurate ML upscaling and forecasting of high-latitude C cycle dynamics.