Rapid, high-fidelity evaluation of multi-component losses in permanent magnet synchronous motor (PMSM) is central to condition monitoring in electric drive systems. To address the prohibitive computational cost of finite element analysis (FEA), the vulnerability of traditional data-driven models to non-physical distortions under sparse samples, and the gradient pathology inherent in joint multi-loss prediction, this paper proposes a surrogate modeling framework integrating high-order physical features and structured mechanistic constraints, namely FMA-PGNN (feature-manifold-augmented physics-guided neural network). Specifically, the framework explicitly extracts nonlinear electromagnetic features via the FMA mechanism. It employs an asymmetric stator-rotor dual-branch architecture embedded with an iron loss operator to achieve mechanistic decoupling, while introducing interval-based soft physical boundaries and dynamic weight scheduling to guarantee global training consistency. Experimental results demonstrate that the proposed method achieves a coefficient of determination exceeding 0.994 and a full-scale relative error strictly below one percent across all loss predictions. Furthermore, it exhibits exceptional extrapolation robustness and physical credibility under unseen operating conditions, providing a highly effective solution for rapid loss evaluation across the entire motor speed domain.

permanent magnet synchronous motor,surrogate modeling,physics-guided neural network,sparse samples,multi-loss prediction