Abstract: Large-eddy simulation (LES), as a key method for simulating complex turbulent phenomena, has demonstrated broad application value in aerospace, combustion propulsion, and atmospheric and oceanic sciences. The core of LES is to use subgrid-scale stress (SGS) models to characterize the nonlinear effects of unresolved small-scale motions on the resolved large-scale flow field, including momentum transport and energy transfer across scales. However, traditional subgrid-scale models are often constrained by single-point flow-field inputs and simplified modeling assumptions, which limit their ability to describe complex multiscale interactions and may lead to large prediction errors. In recent years, data-driven deconvolution methods have developed a new path for turbulent closure modeling by learning or constructing inverse mappings from filtered flow variables to unresolved or unfiltered flow information. This paper reviews the latest progress in data-driven deconvolution subgrid-scale models, focusing on the modeling approaches of deconvolution artificial neural network models (DANN), dynamic iterative approximate deconvolution models (DIAD), and direct deconvolution models (DDM), and verifies their predictive performance through a priori and a posteriori tests. The research shows that reconstructing the unfiltered flow field using data-driven deconvolution methods and then modeling the subgrid-scale unclosed terms can develop highly accurate subgrid-scale models with improved representation of multiscale flow structures. The new generation of deconvolution models not only exhibits very high accuracy in a priori validation (correlation coefficients higher than 0.99, relative error lower than 10%), but also performs excellently in a posteriori tests: without increasing computational costs, its ability to capture turbulent statistics and transient vortex structures is significantly better than traditional models such as — the dynamic Smagorinsky model (DSM) and the dynamic mixed model (DMM), and is very close to the results of direct numerical simulation. These results indicate that data-driven deconvolution methods can construct efficient and highly accurate large-eddy simulation subgrid-scale models for incompressible and compressible turbulence under different filtering conditions.