This paper introduces the Flexible Global Ocean‐Atmosphere‐Land System Model: Grid‐Point Version 3 (FGOALS‐g3) and evaluates its basic performance based on some of its participation in the sixth phase of the Coupled Model Intercomparison Project (CMIP6) experiments. Our results show that many significant improvements have been achieved by FGOALS‐g3 in terms of climatological mean states, variabilities, and long‐term trends. For example, FGOALS‐g3 has a small (−0.015°C/100 yr) climate drift in 700‐yr preindustrial control (piControl) runs and smaller biases in climatological mean variables, such as the land/sea surface temperatures (SSTs) and seasonal soil moisture cycle, compared with its previous version FGOALS‐g2 during the historical period. The characteristics of climate variabilities, for example, Madden‐Julian oscillation (MJO) eastward/westward propagation ratios, spatial patterns of interannual variability of tropical SST anomalies, and relationship between the East Asian Summer Monsoon and El Niño–Southern Oscillation (ENSO), are well captured by FGOALS‐g3. In particular, the cooling trend of globally averaged surface temperature during 1940–1970, which is a challenge for most CMIP3 and CMIP5 models, is well reproduced by FGOALS‐g3 in historical runs. In addition to the external forcing factors recommended by CMIP6, anthropogenic groundwater forcing from 1965 to 2014 was incorporated into the FGOALS‐g3 historical runs.

Four-dimensional variational data assimilation (4DVar) is one of the most promising methods to provide optimal analysis for numerical weather prediction (NWP). Five national NWP centers in the world have successfully applied 4DVar methods in their global NWPs, thanks to the increment method and adjoint technique. However, the application of 4DVar is still limited by the computer resources available at many NWP centers and research institutes. It is essential, therefore, to further reduce the computational cost of 4DVar. Here, an economical approach to implement 4DVar is proposed, using the technique of dimensionreduced projection (DRP), which is called “DRP-4DVar.” The proposed approach is based on dimension reduction using an ensemble of historical samples to define a subspace. It directly obtains an optimal solution in the reduced space by fitting observations with historical time series generated by the model to form consistent forecast states, and therefore does not require implementation of the adjoint of tangent linear approximation. To evaluate the performance of the DRP-4DVar on assimilating different types of mesoscale observations, some observing system simulation experiments are conducted using MM5 and a comparison is made between adjoint-based 4DVar and DRP-4DVar using a 6-hour assimilation window.

A novel full‐field initialization strategy based on the dimension‐reduced projection four‐dimensional variational data assimilation (DRP‐4DVar) is proposed to alleviate the well‐known initial shock occurring in the early years of decadal predictions. It generates consistent initial conditions, which best fit the monthly mean oceanic analysis data along the coupled model trajectory in 1 month windows. Three indices to measure the initial shock intensity are also proposed. Results indicate that this method does reduce the initial shock in decadal predictions by Flexible Global Ocean‐Atmosphere‐Land System model, Grid‐point version 2 (FGOALS‐g2) compared with the three‐dimensional variational data assimilation‐based nudging full‐field initialization for the same model and is comparable to or even better than the different initialization strategies for other fifth phase of the Coupled Model Intercomparison Project (CMIP5) models. Better hindcasts of global mean surface air temperature anomalies can be obtained than in other FGOALS‐g2 experiments. Due to the good model response to external forcing and the reduction of initial shock, higher decadal prediction skill is achieved than in other CMIP5 models.

Marked by large interannual variability, East Asian summer monsoon (EASM) rainfall has profound socio‐economic impacts through its dominant influence on floods and droughts. Improving predictions of the interannual variations of EASM rainfall has important implications for over 20% of the world's population. While coupled modeling systems have demonstrated some prediction skill related to the El Niño Southern Oscillation with remote influence on EASM rainfall, the impact of soil moisture has heretofore not been systematically investigated. Using a weakly coupled data assimilation (WCDA) system to constrain the soil moisture and soil temperature in a coupled climate model with a global land data assimilation product, this study demonstrates significant improvements in simulating the interannual variations of EASM rainfall, capturing the notable shift to a “wetter‐South‐drier‐North” rainfall pattern in China in the early 1990s. Hindcast simulations initialized with the well‐balanced states from a coupled simulation with WCDA also show significant multi‐year rainfall prediction skill over East China and Tibetan Plateau. Improvements in predicting the EASM rainfall are attributed to the strong land‐atmosphere coupling in large areas over China, which allows improved predictions of soil moisture to influence precipitation through soil moisture‐precipitation feedback, and the effects of land anomalies on the EASM circulation. This study highlights the significant contribution of land to the interannual predictability of EASM rainfall, with a great potential to advance skillful interannual predictions of benefit to the large populations influenced by the annual whiplash of the summer monsoon rain.

Ocean waves, especially extreme waves, are vital for air‐sea interaction and shipping. However, current wave models still have significant biases. Based on a numerical wave model and a deep learning model, a BU‐Net by adding batch normalization layers to a U‐Net, we accurately predict the significant wave height (SWH) of the Northwest Pacific Ocean. For each day in 2017–2021, we conducted a 3‐day hindcast experiment using WAVEWATCH3 (WW3) to obtain the SWH forecasts at lead times of 24, 48, and 72 hr, forced by GFS real‐time forecast surface winds. After using BU‐Net, the mean Root Mean Squared Errors (RMSEs) of the SWH forecast from WW3 at lead times of 24, 48, and 72 hr are reduced by 40%, 38%, and 30%, respectively. During typhoon passages, the drop percentages of RMSEs all exceed 20% for three lead times. Therefore, combining numerical models and deep learning is very promising in wave forecasting.