Problem-independent machine learning (PIML) has shown strong potential for accelerating topology optimization by efficiently evaluating numerical shape functions[1]. However, most of PIML-based approaches are restricted to structured grids, limiting their applicability to complex geometries and high-resolution design domains. This work extends the PIML framework to structures with complex geometries by introducing isoparametric elements. To overcome the issue of structured grids, we replace them with isoparametric elements and construct an artificial neural network (ANN) to construct the ML model[2]. The offline-trained ANN incorporates the physical significance of the stiffness matrix to ensure problem independence. The coordinates and material densities of the isoparametric elements serve as inputs for directly predicting the numerical shape functions of the substructures. During optimization, the pre-trained model is employed directly without the need for retraining, enabling the rapid evaluation of shape functions for extended multiscale isoparametric elements. Combined with the constraint of displacement governing equations, the proposed method significantly improves computational efficiency while maintaining high accuracy. Numerical results demonstrate that the finite element analysis efficiency is enhanced by approximately two orders of magnitude, and the computational cost of a single topology optimization step is reduced by a factor of up to 6.7. Furthermore, the method enables high-resolution topology optimization in complex design domains. The proposed approach provides an efficient and scalable solution for topology optimization involving complex geometries and large-scale computational mechanics problems.