With the high-quality development of topology optimization techniques and the rapid proliferation of artificial intelligence, increasing attention has been paid to high-resolution topological designs, and machine learning has been widely leveraged to accelerate structural analysis and design. However, most existing studies offer limited flexibility in handling structural resolution, and the adopted learning-based acceleration schemes often lack generality and robust transferability. In this work, a flexible variable-node substructure element is developed within the classical substructure framework. By exploiting the capability of discrete-variable topology optimization[1] to represent crisp, well-defined topologies, the proposed approach combines the high accuracy of exact substructure with the high efficiency of boundary linear-interpolation substructure. Furthermore, a problem-independent machine learning[2] (PIML) strategy is introduced to further reduce the runtime of finite element analyses by avoiding expensive online computations. Specifically, during the analysis–optimization process, the coarse-grid nodal states are adaptively determined according to the evolving material distribution and partial mechanical response information, so that computational resources are concentrated on regions of interest. The proposed method thereby improves solution accuracy effectively without a significant increase in computational cost, and its effectiveness is demonstrated through representative numerical examples.