arXiv paper targets gap between a priori and a posteriori accuracy in neural network subgrid stress models

A revised arXiv preprint examines why neural network subgrid stress models perform well in a priori tests but deteriorate in a posteriori large eddy simulations. The authors propose approaches to reduce this discrepancy so that model evaluation better reflects real simulation behavior. The work sits in computational fluid dynamics research rather than commercial AI deployment.