Selecting high-performance object-relational mappings (ORMs) requires balancing competing objectives such as insert latency, query latency, and memory usage. In practice, ORM frameworks typically apply fixed heuristics and emit a single schema, while prior learning-based approaches often reduce selection to binary Pareto-membership classification, thereby limiting fine-grained tradeoff reasoning. We present \textbf{Y-Map}, a neural–symbolic framework that predicts \emph{continuous} performance metrics for ORM schema candidates without executing database workloads during inference. Y-Map constructs supervised training data offline via specification-driven schema synthesis and benchmarking, then learns a regression model over fused schema representations that combine interpretable structural descriptors with embeddings from GraphCodeBERT, CodeT5+, and LLaMA. Predicted vectors $\langle \hat{T}{\text{ins}}, \hat{T}{\text{qry}}, \hat{M} \rangle$ enable benchmark-free Pareto-style selection, supporting exhaustive scoring for small spaces and budgeted search for larger spaces. Across nine benchmark object models, Y-Map improves tradeoff quality and reduces inference cost relative to a strong learning-based baseline, demonstrating the promise of combining symbolic validity guarantees with learned performance prediction for practical performance-aware schema design.