Using process-based models to enhance observations from distributed drought experiments.

Distributed experimental networks (DENs) provide unprecedented opportunities to quantify how ecosystem responses to global change across gradients. Process-based models have been previously used together with such manipulative experiments to identify knowledge gaps in models. However, experiments suffer from their own limitations in terms of what, and when, measurements can be taken. Models can offer a powerful tool to provide additional information and key insight into different stages of such experiments. Here we demonstrate that DENs and process-based models can serve a bi-directional diagnostic role: testing model fidelity while revealing where experimental design may systematically bias inference

We simulated ‘International Drought Experiment’ (IDE) drought treatments at grassland site lasting over two growing seasons using QUINCY, a land-surface model of coupled C-N-P cycling, and compared simulated aboveground net primary productivity (ANPP) responses with experimental observations. The model can reproduce the observed relationship between ANPP drought response and drought severity, with overlapping slope confidence intervals (experimental: 0.60 [0.30-0.90]; simulated: 0.79 [0.40-1.18]). Mean simulated ANPP reductions (36%) aligned with IDE synthesis estimates (21-38%), although site by site comparison shows a poorer fit.

Beyond this simple model-data comparison, we can use the model to explore aspects of ecosystem behaviour that cannot easily be measured. We performed model simulations over a range of drought intensities for each site and show that multiple sites exhibited a threshold behaviour – abrupt productivity declines over narrow exclusion ranges. Second, 63% of simulated sites displayed growing season shifts (≥1 month) during drought, with 29% in the first year. The interplay between these two mechanisms – threshold like responses and phenological shifts – produced a critical effect: depending on harvest timing, the magnitude and in some cases the sign of apparent ANPP changes varied substantially.

On the basic model evaluation side, site-level mismatches reflect potential structural constraints in the model (coarse plant functional types, absent competition dynamics producing threshold-like responses). Critically, and in addition to simple model evaluation, the widespread prevalence of growing-season shifts (63% of sites) demonstrates that point sampling could systematically bias inference even in well-designed, standardized experiments – a constraint that cannot be detected from the experimental data alone. This demonstrates that models can enhance the information from manipulative experiments and could either be used post-hoc, as in our study, to add insights to experimental data or prior to experiments to guide design and sampling regimes.