Causal Drivers of Land‐Atmosphere Carbon Fluxes From Machine Learning Models and Data.

Interactions among atmospheric, root‐soil, and vegetation processes drive carbon dioxide fluxes (Fc) from land to atmosphere. Eddy covariance measurements are commonly used to measure Fc at sub‐daily timescales and validate process‐based and data‐driven models. However, these validations do not reveal process interactions, thresholds, and key differences in how models replicate them. We use information theory‐based measures to explore multivariate information flow pathways from forcing data to observed and modeled hourly Fc, using flux tower data sets in the Midwestern U.S. in intensively managed corn‐soybean landscapes. We compare multiple linear regressions, long‐short term memory, and random forests (RF), and evaluate how different model structures use information from combinations of sources to predict Fc. We extend a framework for model predictive and functional performance, which examines a suite of dependencies from all forcing variables to the observed or modeled target. Of the three model types, RF exhibited the highest functional and predictive performance, correctly capturing strong dependencies between radiation and temperature variables with Fc. Regionally trained models demonstrate lower predictive but higher functional performance compared to site‐specific models, suggesting superior reproduction of observed relationships at the expense of predictive accuracy. This study shows that some metrics of predictive performance encapsulate functional behaviors better than others, highlighting the need for multiple metrics of both types. This study improves our understanding of carbon fluxes in an intensively managed landscape, and more generally provides insight into how model structures and forcing variables translate to interactions that are well versus poorly captured in models. Plain Language Summary: In an agricultural landscape, exchanges of carbon dioxide between the land and atmosphere occur due to photosynthesis and respiration and depend on weather, soil, and vegetation conditions. Traditional model performance metrics focus on the relationship between observed and modeled outputs, while functional performance considers the relationships between interacting inputs and outputs. We compare several performance measures for three different machine learning (ML) models that simulate sub‐daily carbon fluxes. We look at how drivers such as solar radiation, soil moisture, temperature, humidity, and rainfall provide information to carbon fluxes, and whether different ML models also capture these interactions. In other words: Air, soil, and plants drive carbon's upward path,Models are detectives, interpreting their math.With information theory, we map data's travel courses,To see how models find or miss carbon's causal sources. Key Points: Information theory measures describe individual and joint causal relationships in observed versus modeled vertical carbon dioxide fluxesThree machine learning models overestimate unique information from sources at the expense of synergistic, or joint informationRegionally trained models have improved functional but not predictive performances, indicating a trade‐off [ABSTRACT FROM AUTHOR]