This letter proposes a novel neuro-coupling matrix (neuro-CM) technique for parametric modeling of microwave filters. It is the first time to combine coupling matrix (CM) and neural networks calculating intermediate variables to learn the relationship between geometrical parameters and the electromagnetic (EM) response of microwave filters. A novel center-out optimization method is proposed to extract the CM parameters as training data more effectively, which provides much more continuous intermediate parameters than vector fitting. Compared with the existing neuro-transfer function (neuro-TF) method, the proposed neuro-CM method can achieve better accuracy with a wider geometrical range. The effectiveness of the proposed method is verified through two microwave filter examples.