Prediction in trait-based ecology: global simulations of specific leaf area using a trait-based dynamic vegetation model

Predicting plant functional traits is crucial because traits underpin ecosystem processes. Statistical, machine-learning, and optimality approaches have produced global plant trait predictions. However, the utility of trait-based vegetation models remains unexplored at this scale. We use aDGVM2, an eco-evolutionary trait- and individual-based dynamic global vegetation model (DGVM). aDGVM2 simulates community assembly driven by natural selection. Simulated specific leaf area (SLA) is an emergent outcome of community assembly driven by abiotic conditions and plant competition. We examine (1) how well aDGVM2 simulates global SLA compared to trait data and (2) what drives large differences. Compared to GBIF-derived SLA data, aDGVM2 displays a mean SLA difference of -2.9 m2 kg-1 (GBIF range ca. 4-35 m2 kg-1), a root mean square error (RMSE) of 7.26, and a normalised mean absolute error (nMAE) of 26.56 %. Published approaches showed differences of -4.83 to 2.71 (mean), 4.42-6.81 (RMSE), and 11.64 %-25.23 % (nMAE). aDGVM2 mean differences are comparable with published predictions, while RMSEs and nMAEs are higher. Large aDGVM2 mismatches occur where deciduous and evergreen abundances are inaccurate. Correcting abundance mismatches using observed leaf phenology abundances strongly reduces differences (mean:-0.11 to 0.43, RMSE: 5.86-5.91, nMAE: 15.10 %-19.61 %), yielding the smallest mean difference and second-smallest nMAE of all products compared. Thus, an eco-evolutionary, process-based approach can predict global SLA values as well as the best available products once leaf phenological abundances are accurate. Improving the simulation of these abundances is therefore a viable route towards a process-based understanding of the global distribution of SLA and functional trait diversity under contemporary and future climate conditions.