Structural topology optimization (TO) has become a cornerstone in industrial design, enabling efficient material distribution under specified constraints. Traditional approaches such as the SIMP method face significant limitations when attempting to generate trabecular-like structures, due to strong dependencies between geometric features and mesh resolution—ultimately resulting in high computational costs. To address this, several multiscale strategies have been developed to reduce computation time . Among them, the 2-Level TO technique stands out by first solving the TO problem on a coarse grid, then refining the solution through localized optimizations in each subdomain, while preserving global structural continuity. In contrast to other existing approaches, the 2-Level TO method does not assume scale separation, which is indeed inapplicable in this context. However, in increasingly demanding industrial applications—and even in certain biomedical contexts—there is a growing need to further reduce computational effort. In this regard, machine learning (ML) approaches, particularly one-shot methods , have proven effective at accelerating TO workflows. Nonetheless, these methods often suffer from issues of structural inconsistency, typically requiring additional SIMP iterations for correction. To overcome these limitations, we propose a data-driven framework that integrates one-shot prediction techniques within the 2-Level TO paradigm. Our approach emphasizes the definition of a distance metric in the parameter space to provide good initial solutions. The method also includes a contextualized dataset generation method based on a linear manifold learning strategy, thus alleviating the computational burden of exploring the high-dimensional parametric space. Consequently, the use of this metric provides a structurally sound topology, which may be adapted to the actual loads with a reduced number of SIMP iterations, achieving reductions of up to 75% in computational cost. AKNOWLEDGEMENTS The authors gratefully acknowledge the financial support of: Grant MCIN/AEI/10.13039/501100011033 funded by ”Ministerio de Ciencia e Innovación”. Grant PID2022-141512NB-I00 funded by ”Fondo Europeo de Desarrollo Regional -FEDER-”