Abstract

Divergence in land carbon cycle simulation is persistent and widespread. Regardless of model intercomparison project, results from individual models diverge significantly from each other and, in consequence, from reference datasets. Here we link model spread to structure using a 15-member ensemble of land surface models from the Multi-scale synthesis and Terrestrial Model Intercomparison Project (MsTMIP) as a test case. Our analysis uses functional benchmarks and model structure as predicted by model skill in a machine learning framework to isolate discrete aspects of model structure associated with divergence. We also quantify how initial conditions prejudice present-day model outcomes after centennial-scale transient simulations. Overall, the functional benchmark and machine learning exercises emphasize the importance of ecosystem structure in correctly simulating carbon and water cycling, highlight uncertainties in the structure of carbon pools, and advise against hard parametric limits on ecosystem function. We also find that initial conditions explain 90% of variation in global satellite-era values—initial conditions largely predetermine transient endpoints, historical environmental change notwithstanding. As MsTMIP prescribes forcing data and spin-up protocol, the range in initial conditions and high levels of predetermination are also structural. Our results suggest that methodological tools linking divergence to discrete aspects of model structure would complement current community best practices in model development.