Compared to the traditional periodic lattice structures, conformal lattice structures offer superior shape-adaptivity and a broader performance design space [1,2]. However, their spatially varying microstructures violate the periodic assumption, rendering conventional homogenization methods ineffective. To resolve this conflict, our work proposes a multiscale computational framework, that couples a homogenization method based on the local periodicity assumption as microsolver, with the heterogeneous multiscale method as macrosolver [3], for the multiscale computation of aperiodic lattice structures. Within this multiscale framework, the weak form calculations of each integration domain derive constitutive constants to determine the effective mechanical properties distribution of the aperiodic structure. Moreover, to ensure prediction efficiency, the efficient surrogate model is constructed by multi-output Gauss process regression combined with a PCA-based dimension reduction method. With the active learning [4] in the latent constitutive space, the surrogate model compressively utilizes statistics posterior performance for the selection of the most informative sampling points for training, thereby achieving accurate prediction for the anisotropic constitutive performances utilizing possibly small scale dataset. Additionally, by explicitly incorporating similarity transformations, comprising rigid-body rotation, mirror symmetry and scaling, the model further reduces the sampling scale for the input dataset [5]. This work addresses a key scientific question in multiscale computation: how to perform exact multiscale analysis of conformal aperiodic lattice structures to predict their effective mechanical responses, with as few homogenization computations as possible