Understanding and reducing the uncertainties of land surface energy flux partitioning within CMIP6 land models

•CMIP6 models largely overestimated the land surface evaporative fraction (defined as LE/(LE + H)).•Surface forcing variables e.g., vapor pressure deficit dominated the model EF uncertainties.•Accounting the forcing biases could largely improve CMIP6 model ensembles.•ML-based asurrogate model could further improve the CMIP6 modeled EF through parameterization.•CMIP6 model structure imperfection prevailed at evergreen broadleaf forest ecosystems.