New Energy Vehicles (NEVs) drive China's automotive market. However, calibrating thermal management for the battery, motor, and inverter is a major bottleneck, especially as development cycles shrink to 24 months. High-fidelity thermal-fluid models are too slow for repeated calibration and fixed-step real-time execution. This paper presents an end-to-end industrial workflow: 1) Build high-fidelity models using 3D data; 2) Train reduced-order models (ROMs), implemented here as nonintrusive data-driven surrogate models, from simulation data; 3) Export the trained surrogates as FMI-compliant co-simulation FMUs for NI VeriStand real-time testing. The ROM tool proves efficient for repeated calibration runs; in this paper, ROM refers to a non-intrusive datadriven surrogate model exported as an FMI-compliant cosimulation FMU, rather than a projection-based or equation-level model-order-reduction method. The reported speedup is interpreted as repeated-simulation acceleration, with data generation and training treated as one-time amortized costs. For the reported battery validation cases, temperature RMSE was evaluated against experimental or CAE reference data and normalized by the initial temperature of the battery heat-generation operating condition. The calculation results show that the single-cell RMSE is within 5%, the battery-pack RMSE is within 10%, and the absolute temperature error is within 3 K. For a representative 10-module battery-pack cooling case, the 250 s closed-loop simulation time was reduced from 2200 s to 44 s after surrogate-FMU replacement, and the calibrated PI response met the control targets.