Accurate prediction of compressor performance maps and speedlines using high-fidelity CFD remains computationally expensive due to the strong nonlinearity of turbomachinery flows, sensitivity to operating conditions, and the presence of critical phenomena such as flow separation, rotating stall inception, surge, and choke. These challenges are particularly pronounced near the limits of the operating envelope, where small variations in boundary conditions can lead to abrupt changes in flow behavior and performance. This work presents a physics-constrained, data-driven modeling approach for compressor speedline prediction based on a surrogate modeling framework. The proposed method is designed to efficiently learn dominant flow features while preserving physically meaningful relationships between operating parameters and aerodynamic performance. Special emphasis is placed on modeling extended operating ranges and accurately capturing behavior near surge and choke margins, where traditional surrogate models often exhibit instability and reduced accuracy. To assess robustness in these highly nonlinear regimes, different model architectures and training configurations are evaluated to determine their suitability and applicability for compressor speedline prediction. Ensemble learning concepts are explored to improve robustness and reduce predictive uncertainty, which is critical for reliable performance assessment in regions characterized by strong flow instabilities. In addition, the proposed machine learning approach is benchmarked against a reduced order POD (Proper Orthogonal Decomposition) based Gaussian Process Regression method to assess the relative strengths of physics aware learning and classical reduced order modeling techniques in representing complex turbomachinery flow dynamics.