Reducing the structural uncertainty of global lake evaporation rate projection

Evaporation is a key component of freshwater loss of lakes. Policy makers need reliable evaporation projection for adaptive allocation of water resources. However, large uncertainty exists in global lake evaporation rate (E) projection. When scenario is fixed, the uncertainty is mainly arisen from climate model uncertainty, that is the choice of Earth system model (ESM) outputs to drive lake model. However, the relative contribution of ESMs structural uncertainty is still unclear. Furthermore, there is no physics-informed method to reduce structural uncertainty. A primary reason is that multi-model ensemble projections of lake E in online mode are still absent. To address the shortcoming, we firstly combined Community Earth System Model 2 (CESM2), the only one with lake E projections under SSP370 in CMIP6, with automatic machine learning algorithm to establish a global lake E emulator. The emulator “solves” the lake E statistically with high efficiency instead of numerically. The dynamic interactions between lake and atmosphere are also preserved in the emulator by training with the CESM2 Large Ensemble (LENS2). The emulator can produce global online multi-model projections of lake E under SSP370 scenario with 30 ESM atmospheric forcing variables. Then, the structural uncertainty is calculated as standard deviation among multiple ESMs. At last, the emergent constraints for lake E structural uncertainty were established in different climate zones and at the global scale. The results show that structural uncertainty is the largest for tropical lakes. VPD is an optimal variable used for emergent constraints. After emergent constraints, Lake E in tropical climate will increase a little faster with reduced uncertainty (~23%). This study can provide theory support for enhancing credibility of future lake water storage projection, also show the direction for improving lake processes simulation in next generation of ESMs.