Abstract <p>Multiscale structural optimization is essential for achieving lightweight and high-performance designs in advanced engineering applications, yet direct topology optimization of multiscale structures remains computationally prohibitive. Moreover, conventional compliance-based cellular designs often suffer from severe stress concentrations, especially near sharp features and re-entrant corners. To address these challenges, this study proposes a novel data-driven two-scale hierarchical aggregation framework for stress minimization in graded cellular structures. At the macroscale, a topological description function is employed as the design variable to control material distribution. At the micro-scale, microstructure is defined using a level set method by a microscopic design variable. The background mesh is generated by mmg, which is an open-source software for bidimensional and tridimensional surface and volume remeshing. An offline, problem-independent database is constructed via numerical homogenization and interpolation, capturing the relationships between microstructural parameters and macro-element properties. During optimization, microscopic stresses are aggregated into macroscopic element stresses, which are then condensed into a global p-norm stress metric. The entire topology optimization is conducted on the macroscale, leveraging the precomputed database to significantly reduce computational cost. After optimization, full-scale geometric reconstruction is performed using interpolation techniques. Numerical results demonstrate that the proposed framework effectively reduces stress concentrations and achieves superior stress performance compared with conventional compliance-based topology optimization.</p> Graphical abstract <p></p>