DRL-EnVar: an adaptive hybrid ensemble–variational data assimilation method based on deep reinforcement learning

• Proposes DRL-EnVar, a deep reinforcement learning-based method for adaptive hybrid ensemble–variational data assimilation, dynamically optimizing hybrid weights. • A novel cyclic convolution module extracts abstract features from data to improve the estimation of background error covariance. • Outperforms traditional EnKF and hybrid covariance DA methods, especially under sparse observations and transitional weather regimes, with competitive or superior accuracy at lower computational cost. • Can be flexibly integrated into both 3DVar and 4DVar frameworks, offering a novel approach for improving forecast skill during transitional weather states.