CLM5 experimental simulations

In this study, the CLM was run in the offline mode. To achieve carbon, water and energy steady state at the start of the simulations⁶², CLM5 was run for 400 years in the “spin‐up” mode with accelerated decomposition rates of soil organic matters, followed by an additional 800 years’ run in the “normal mode”. By the end of the spin‐up, the total ecosystem C was drifting by less than 0.02 Pg C year^–1 at global scale, much lower than the 0.1 Pg C yr^-1 threshold recommended for the C4MIP by Jones et al.⁶³. Then, experimental simulations were conducted using CLM5 for varying wind speed scenarios for the period of 1983 – 2014, including:i) Control run: The simulation was forced by the default air forcing dataset of CLM5 – CRUNCEP version 7 meteorological data (6-hourly data; https://rda.ucar.edu/datasets), except for wind speed, which was replaced by data from the CRU JRA version 2.2 dataset, because the CRUNCEP7 wind speed revealed widespread increasing trend of terrestrial wind speed during 1983 – 2014, inconsistent with observations; in contrast, CRU JRA2.2 wind speed showed overall decreasing trend of terrestrial wind speed during the same period. All air forcing data had a temporal resolution of 6-hourly and a spatial resolution of 0.5° × 0.5°, and the simulation was conducted at 0.9° × 1.25° resolution. The atmospheric CO₂ concentration in the simulation was set to a constant value of 367 ppm (default value in the model) to exclude effects of the rising CO₂ concentration on GPP.ii) Sensitivity run: All model settings and drivers were the same as the control run, except that the wind speed was uniformly increased / decreased by 1% and 8% respectively throughout the simulation period. The extracted declining trends in spatially-averaged terrestrial wind speed during 1983–2014 were -1.48% and -1.95% per decade from CRU dataset and CESM2 simulation, respectively (Fig. 1a). Previous studies reported an averaged terrestrial wind speed decline of up to 8% over the past three decades³. Adjustments of 1% and 8% were determined to cover the range of the terrestrial wind speed changes. We haven’t further manipulated wind speed within the range of 1%– 8%, because the responses of GPP to wind speed changes (ΔGPP) from -8% to +8% are essentially linear (Fig. S12).iii) Detrended wind speed run: All model settings were the same as the control run, except that the long-term trend in wind speed was eliminated. The singular spectrum analysis⁶⁴ was used to extract and remove the long-term trend in wind speed, and the extraction was carried out on a 5-day scale, instead of 6-hour or daily scale, to deal with the requirements of substantial computational resources. Specifically, trend extraction was implemented using function ‘trenddecomp’ in MATLAB for each terrestrial grid.