Imitation or identification: limitations of deep learning in extrapolating to future climate-carbon cycle change

A new frontier in Earth system modeling has emerged, with machine learning (ML) and deep learning (DL) holding great promise to overcome long-standing challenges in future climate change projections. Yet, a fundamental question persists: Can these DL models reliably extrapolate to unseen climatic conditions? Here, we test whether various DL architectures successfully extrapolate both spatially and temporally in a controlled setting that mimics real-world complexity. Focusing on land CO _2 uptake dynamics as a test case—a system substantially impacted by rising CO _2 levels and climate change—we train a diverse array of DL models using historical time-series simulated data from 1850 to the present and put them to the test for a future climate change scenario. While these DL models excel in spatial cross-validation (normalized Nash–Sutcliffe efficiency (NNSE) greater than 0.9) and exhibit decent performance across different continents (NNSE $ \gt $ 0.7), most show virtually no skill when predicting CO _2 uptake under future conditions (NNSE $ \lt $ 0.6). Notably, the DL models largely overestimate future land carbon uptake, driven by an exaggerated sensitivity to CO _2 that results in a physically implausible shift in the seasonal cycle. Explainable ML techniques reveal that strong signals, e.g. the effect of solar radiation, are accurately learned; however, nuanced long-term effects critical for climate projections, e.g. CO _2 fertilization, are not captured. These findings underscore the need for caution in applying ML for extrapolation and highlight the necessity of developing techniques that align with expected extrapolation behaviors of Earth system processes.